From 02f33767c113f12426585f6fd86b91d255fc9afe Mon Sep 17 00:00:00 2001
From: Abhijeet Prasad <abhijeet@braintrustdata.com>
Date: Wed, 23 Sep 2026 11:22:53 -0400
Subject: [PATCH 1/4] fix(anthropic): capture usage and refusal metadata

Preserve fast-mode and fallback-credit usage fields so spans retain the
billing context Anthropic reports. Record compaction and context-management
request settings, and include refusal stop details to show why a response was
refused.

Refs #791
---
 .../integrations/anthropic/_utils.py          |  2 +
 .../integrations/anthropic/test_anthropic.py  | 40 +++++++++++++++++++
 .../integrations/anthropic/tracing.py         |  3 ++
 3 files changed, 45 insertions(+)

diff --git a/py/src/braintrust/integrations/anthropic/_utils.py b/py/src/braintrust/integrations/anthropic/_utils.py
index e6c38273..fbd264d9 100644
--- a/py/src/braintrust/integrations/anthropic/_utils.py
+++ b/py/src/braintrust/integrations/anthropic/_utils.py
@@ -32,6 +32,8 @@ def __getattr__(self, name: str) -> Any:
     {
         "service_tier",
         "inference_geo",
+        "speed",
+        "fallback_credit",
     }
 )
 
diff --git a/py/src/braintrust/integrations/anthropic/test_anthropic.py b/py/src/braintrust/integrations/anthropic/test_anthropic.py
index 55b531dd..0ae95792 100644
--- a/py/src/braintrust/integrations/anthropic/test_anthropic.py
+++ b/py/src/braintrust/integrations/anthropic/test_anthropic.py
@@ -268,6 +268,21 @@ def test_get_metadata_from_kwargs_includes_structured_output_params():
     }
 
 
+def test_get_metadata_from_kwargs_includes_compaction_and_context_management():
+    metadata = _get_metadata_from_kwargs(
+        {
+            "compaction": [{"type": "enabled"}],
+            "context_management": {"edits": [{"type": "clear_tool_uses_20250919"}]},
+        }
+    )
+
+    assert metadata == {
+        "provider": "anthropic",
+        "compaction": [{"type": "enabled"}],
+        "context_management": {"edits": [{"type": "clear_tool_uses_20250919"}]},
+    }
+
+
 def test_log_message_to_span_includes_stop_reason_and_stop_sequence():
     span = unittest.mock.MagicMock()
     message = SimpleNamespace(
@@ -276,6 +291,7 @@ def test_log_message_to_span_includes_stop_reason_and_stop_sequence():
         model=MODEL,
         stop_reason="stop_sequence",
         stop_sequence="DONE",
+        stop_details=None,
         usage={
             "input_tokens": 11,
             "output_tokens": 7,
@@ -312,6 +328,26 @@ def test_log_message_to_span_includes_stop_reason_and_stop_sequence():
     )
 
 
+def test_log_message_to_span_includes_refusal_stop_details():
+    span = unittest.mock.MagicMock()
+    message = SimpleNamespace(
+        role="assistant",
+        content=[],
+        model=MODEL,
+        stop_reason="refusal",
+        stop_sequence=None,
+        stop_details={"category": "cyber", "explanation": "unsafe request"},
+        usage={},
+    )
+
+    _log_message_to_span(message, span)
+
+    assert span.log.call_args.kwargs["output"]["stop_details"] == {
+        "category": "cyber",
+        "explanation": "unsafe request",
+    }
+
+
 def test_extract_anthropic_usage_includes_server_tool_use_metrics_from_objects():
     usage = SimpleNamespace(
         input_tokens=11,
@@ -368,6 +404,8 @@ def to_dict(self):
                 }
             ),
             "service_tier": "standard",
+            "speed": "fast",
+            "fallback_credit": 0.25,
         }
     )
 
@@ -385,6 +423,8 @@ def to_dict(self):
     }
     assert metadata == {
         "usage_service_tier": "standard",
+        "usage_speed": "fast",
+        "usage_fallback_credit": 0.25,
     }
 
 
diff --git a/py/src/braintrust/integrations/anthropic/tracing.py b/py/src/braintrust/integrations/anthropic/tracing.py
index d4030e90..01a5552c 100644
--- a/py/src/braintrust/integrations/anthropic/tracing.py
+++ b/py/src/braintrust/integrations/anthropic/tracing.py
@@ -72,6 +72,8 @@ def _accumulate_event_accepts_json_bufs() -> bool:
     "thinking",
     "output_config",
     "output_format",
+    "compaction",
+    "context_management",
 )
 
 
@@ -1512,6 +1514,7 @@ def _message_output(message, *, include_parsed_output: bool = False):
             "model": getattr(message, "model", None),
             "stop_reason": getattr(message, "stop_reason", None),
             "stop_sequence": getattr(message, "stop_sequence", None),
+            "stop_details": getattr(message, "stop_details", None),
         }.items()
         if v is not None
     }

From 422b42e5d80b41a5b9fab3aa7e35eb5c78ef1efc Mon Sep 17 00:00:00 2001
From: Abhijeet Prasad <abhijeet@braintrustdata.com>
Date: Wed, 23 Sep 2026 12:12:42 -0400
Subject: [PATCH 2/4] commit message

---
 .agents/skills/commit-message/SKILL.md        | 169 +++---------------
 .../a28062d365df/manifest.json                |  27 +++
 .../a28062d365df/spec.json                    |  12 ++
 .../a28062d365df/state.json                   |  17 ++
 .../025f8d50cb22/manifest.json                |  27 +++
 .../025f8d50cb22/spec.json                    |  12 ++
 .../025f8d50cb22/state.json                   |  17 ++
 7 files changed, 133 insertions(+), 148 deletions(-)
 create mode 100644 bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json
 create mode 100644 bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json
 create mode 100644 bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json
 create mode 100644 bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json
 create mode 100644 bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json
 create mode 100644 bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json

diff --git a/.agents/skills/commit-message/SKILL.md b/.agents/skills/commit-message/SKILL.md
index 6a44ebf0..23717f27 100644
--- a/.agents/skills/commit-message/SKILL.md
+++ b/.agents/skills/commit-message/SKILL.md
@@ -1,169 +1,42 @@
 ---
 name: commit-message
-description: Generates a meaningful conventional commit message (title + body) from the current diff and conversation context. Use when asked to write, suggest, or generate a commit message for the current changes.
+description: Draft a conventional commit message from the current changes and conversation. Use when asked to write, suggest, or generate a commit message; do not commit unless asked.
 ---
 
 # Commit Message Generator
 
-Generate a conventional commit message based on the staged/unstaged diff and the surrounding conversation.
+Write a concise, accurate commit message that captures the intent of the current changes. Use the conversation for motivation and constraints; use the diff as evidence. Do not claim behavior or test results the changes do not support.
 
-## Conventional Commits Format
+## Gather context
 
-```
-<type>(<scope>): <short summary>
-
-<body>
-
-<footer>
-```
-
-**Types:**
-- `feat` — new feature
-- `fix` — bug fix
-- `ref` — code restructuring without behavior change (prefer `ref` over `refactor`)
-- `perf` — performance improvement
-- `test` — adding or updating tests
-- `docs` — documentation only
-- `chore` — build system, tooling, dependencies, config
-- `ci` — CI/CD pipeline changes
-- `style` — formatting, whitespace (no logic change)
-- `revert` — reverts a previous commit
-
-**Rules:**
-- Subject line ≤ 72 characters, lowercase, no trailing period
-- Imperative mood: "add feature" not "added feature"
-- Body wraps at 100 characters, explains *what* and *why* (not *how*)
-- Footer: `BREAKING CHANGE: ...` or `Fixes #<issue>` / `Closes #<issue>` if applicable
-- Scope is optional but encouraged when it adds clarity (e.g., `feat(auth):`, `fix(api):`)
-- Omit body if the subject line fully communicates the intent
-
-## Writing Style
-
-- Lead with the point. Be direct, technical, collaborative, and lightly casual.
-- Use plain words, contractions where natural, and short paragraphs.
-- Prefer `we` for shared decisions and `I` for genuine opinion or uncertainty. Do not hide uncertainty behind authoritative prose.
-- Ground claims in specifics: name the API, version, behavior, error, test, or file involved.
-- Put code, identifiers, filenames, versions, and literal values in backticks.
-- Give only the context needed to explain **context → change → reason/evidence → consequence or follow-up**.
-- Surface compatibility constraints, tradeoffs, risks, and intentionally deferred work plainly when relevant.
-- Use bullets for multiple distinct changes or findings; avoid unnecessary headings and polished filler.
-- Do not merely restate the diff. Explain why the implementation matters or how behavior changes.
-
-For a substantive change, actively consider including a compact code snippet, before/after example, or ASCII diagram when it explains behavior or data flow more clearly than prose. Keep it focused and omit it when it would be decorative or redundant. For example:
-
-````text
-Before: provider response -> wrapper-specific span
-After:  provider response -> shared integration hook -> normalized span
-````
-
-or:
-
-````python
-# Before
-wrap(client)
-
-# After
-client = wrap_client(client)
-````
-
-## Instructions
-
-1. **Collect the diff** — run `git diff HEAD` (staged + unstaged). If empty, try `git diff --cached` (staged only). If still empty, try `git status --short` and `git log --oneline -3` to understand the trajectory.
-
-2. **Review the conversation** — you already have the conversation history in context. Look for:
-   - Explicit intent from the user ("I'm adding X", "this fixes Y")
-   - Issue or ticket numbers mentioned
-   - Feature names, module names, or domain language used
-   - Any constraints or things the user emphasized
-
-3. **Identify ambiguities** — before generating, check if any of these are unclear:
-   - Is the primary intent a new feature, a fix, a refactor, or something else?
-   - Is there a scope (module/package/component) worth calling out?
-   - Are there breaking changes?
-   - Is there a related issue or ticket number?
-   - Does the diff span multiple unrelated concerns? (should be split into separate commits)
-
-4. **Ask for clarifications if needed** — if the intent is genuinely ambiguous from both the diff and the conversation, ask 1–3 focused questions before generating. Do **not** ask about things already clear from the context.
-
-5. **Generate the commit message** — produce exactly one commit message in a fenced code block:
-   - Pick the most specific `type` that fits
-   - Include a `scope` when it meaningfully narrows the change
-   - Write a crisp subject line in imperative mood
-   - Add a body if the change is non-trivial, explaining the reasoning
-   - Check whether a small before/after snippet or ASCII diagram would make a substantive change easier to review, and include one when it would
-   - Add footer entries for breaking changes or issue references
-   - **CRITICAL when committing:** preserve **real newline characters** in the commit body
-   - **Never** put literal `\n` text inside a quoted `git commit -m "..."` body and assume Git will turn it into line breaks — it will not
-   - If issuing `git commit` yourself, prefer these patterns in this order:
-     1. **Best for multiline bodies:** write the full message to a temp file and use `git commit -F <file>`
-     2. **Good for amendments:** write the full message to a temp file and use `git commit --amend -F <file>`
-     3. multiple `-m` flags, e.g. `git commit -m "subject" -m "first paragraph
-
-second paragraph"`
-     4. ANSI-C quoting, e.g. `git commit -m "subject" -m $'line 1\n\nline 2'`
-   - Prefer temp-file commits by default whenever the body has multiple paragraphs, bullets, or any non-trivial formatting
-   - Before finalizing, sanity-check that `git log -1 --format=medium` shows actual blank lines and wrapped paragraphs, not backslash-n sequences
-
-## Newline safety examples
-
-**Wrong:**
-
-```bash
-git commit -m "docs: add pi guide" -m "line 1\n\nline 2"
-```
-
-This stores the characters `\` and `n` literally in the commit message.
-
-**Correct:**
-
-```bash
-git commit -m "docs: add pi guide" -m $'line 1\n\nline 2'
-```
-
-or
-
-```bash
-git commit -m "docs: add pi guide" -m "line 1
-
-line 2"
-```
+From the repository root, inspect `git status --short`, `git diff HEAD`, and `git diff --cached`. Check untracked paths shown by status when they are part of the requested change. If there is no relevant diff, use recent commit history for context and say when there is not enough information to draft a reliable message.
 
-or use a temp file/editor.
+Review the conversation for the user's stated goal, issue references, and compatibility constraints. If multiple unrelated changes appear, mention that they may need separate commits instead of hiding them in one vague message.
 
-## Default execution preference
+## Format
 
-When the task is not just to suggest a commit message, but to actually run `git commit` or `git commit --amend`:
+Use Conventional Commit form:
 
-1. If the message has a body, prefer a temp file with `-F`
-2. If amending a commit with a body, prefer `git commit --amend -F <file>`
-3. Only use inline `-m` bodies when the formatting is trivially simple and you are certain real newlines will be preserved
-4. After committing, verify with `git log -1 --format=medium`
+```text
+<type>(<scope>): <imperative summary>
 
-This preference exists to avoid malformed commit bodies with literal `\n` sequences.
-
-## Output Format
-
-Present the final message in a fenced code block. If the message contains its own fenced snippet, use a longer outer fence so the full message remains copyable:
+<optional body>
 
+<optional footer>
 ```
-feat(auth): add OAuth2 PKCE flow for CLI login
 
-Replace the device-code flow with PKCE so the CLI can authenticate
-without opening a browser on headless machines. The previous flow
-required interactive browser consent which blocked CI usage.
+- Choose the most specific suitable type: `feat`, `fix`, `ref` (repository convention for behavior-preserving restructuring), `perf`, `test`, `docs`, `chore`, `ci`, `style`, or `revert`.
+- Add a scope only when it meaningfully narrows the change.
+- Keep the subject at 72 characters or fewer, lowercase, imperative, and without a trailing period.
+- Add a body when the subject alone does not explain the purpose or a meaningful behavior change. Keep it focused on what changed and why; wrap prose at about 100 characters.
+- Add `BREAKING CHANGE: ...` or an issue-closing footer only when the diff or conversation supports it.
 
-Closes #342
-```
+Use direct, specific language. Name relevant APIs, versions, errors, or files when they clarify the change. Avoid restating the diff, unsupported claims, decorative examples, and filler.
 
-Then briefly (1–2 sentences) explain the key decision made (type choice, scope, whether a body was needed).
+For substantive changes, prefer a compact code snippet or ASCII diagram in the body when it explains behavior or data flow more clearly than prose. Keep it focused and omit it when it adds no clarity.
 
-If you asked for clarifications and the user answered, incorporate those answers and output the final message immediately — no need to re-ask.
+## Output and commit execution
 
-## Commands
+For a message suggestion, return one copyable message in a fenced code block. Add at most one brief sentence explaining a non-obvious type or scope choice. Ask a focused question only when the intent cannot reasonably be inferred from the diff and conversation.
 
-```bash
-git diff HEAD
-git diff --cached
-git status --short
-git log --oneline -5
-```
+Do not run `git commit` or `git commit --amend` unless the user explicitly asks you to commit. When committing a multiline message, preserve real newline characters: prefer writing the message to a temporary file and using `git commit -F <file>` (or `git commit --amend -F <file>`). Multiple `-m` flags or ANSI-C quoting are acceptable for simple cases. Never assume literal `\n` inside ordinary quotes becomes a newline. After committing, verify the stored message with `git log -1 --format=medium`.
diff --git a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json
new file mode 100644
index 00000000..2054395e
--- /dev/null
+++ b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json
@@ -0,0 +1,27 @@
+{
+  "schema_version": 1,
+  "spec_hash": "a28062d365df1d03f56d9706e01ed761d13d5f5e789e82a903fbe24af390f0f9",
+  "spec": {
+    "schema_version": 1,
+    "object_ref": "project_logs:037d4d7e-7f0f-4064-ad86-af82550839cd",
+    "object_type": "project_logs",
+    "object_name": "037d4d7e-7f0f-4064-ad86-af82550839cd",
+    "direction": "push",
+    "scope": "all",
+    "filter": null,
+    "limit": null,
+    "page_size": 1000,
+    "input_path": "/private/tmp/anthropic-integration-example.jsonl"
+  },
+  "last_run_id": "run-1790179459",
+  "status": "completed",
+  "items_done": 1,
+  "pages_done": 1,
+  "bytes_processed": 1084,
+  "output_path": null,
+  "input_path": "/tmp/anthropic-integration-example.jsonl",
+  "started_at": 1790179459,
+  "updated_at": 1790179460,
+  "completed_at": 1790179460,
+  "message": null
+}
diff --git a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json
new file mode 100644
index 00000000..5f0f1ab9
--- /dev/null
+++ b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json
@@ -0,0 +1,12 @@
+{
+  "schema_version": 1,
+  "object_ref": "project_logs:037d4d7e-7f0f-4064-ad86-af82550839cd",
+  "object_type": "project_logs",
+  "object_name": "037d4d7e-7f0f-4064-ad86-af82550839cd",
+  "direction": "push",
+  "scope": "all",
+  "filter": null,
+  "limit": null,
+  "page_size": 1000,
+  "input_path": "/private/tmp/anthropic-integration-example.jsonl"
+}
diff --git a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json
new file mode 100644
index 00000000..4a504eb8
--- /dev/null
+++ b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json
@@ -0,0 +1,17 @@
+{
+  "schema_version": 1,
+  "run_id": "run-1790179459",
+  "status": "completed",
+  "scope": "all",
+  "limit": null,
+  "page_size": 1000,
+  "source_path": "/tmp/anthropic-integration-example.jsonl",
+  "line_offset": 1,
+  "items_done": 1,
+  "pages_done": 1,
+  "bytes_sent": 1084,
+  "distinct_roots_done": 1,
+  "started_at": 1790179459,
+  "updated_at": 1790179460,
+  "completed_at": 1790179460
+}
diff --git a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json
new file mode 100644
index 00000000..05f3b9bc
--- /dev/null
+++ b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json
@@ -0,0 +1,27 @@
+{
+  "schema_version": 1,
+  "spec_hash": "025f8d50cb2237fecd24ef2dd9911b98a783fd9ffc2611b5d6a15435f3c8c227",
+  "spec": {
+    "schema_version": 1,
+    "object_ref": "project_logs:56f34894-753c-4407-bc0e-09360a5d69ea",
+    "object_type": "project_logs",
+    "object_name": "56f34894-753c-4407-bc0e-09360a5d69ea",
+    "direction": "push",
+    "scope": "all",
+    "filter": null,
+    "limit": null,
+    "page_size": 1000,
+    "input_path": "/private/tmp/anthropic-integration-example.jsonl"
+  },
+  "last_run_id": "run-1790179373",
+  "status": "completed",
+  "items_done": 1,
+  "pages_done": 1,
+  "bytes_processed": 1084,
+  "output_path": null,
+  "input_path": "/tmp/anthropic-integration-example.jsonl",
+  "started_at": 1790179373,
+  "updated_at": 1790179373,
+  "completed_at": 1790179373,
+  "message": null
+}
diff --git a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json
new file mode 100644
index 00000000..eef0f45f
--- /dev/null
+++ b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json
@@ -0,0 +1,12 @@
+{
+  "schema_version": 1,
+  "object_ref": "project_logs:56f34894-753c-4407-bc0e-09360a5d69ea",
+  "object_type": "project_logs",
+  "object_name": "56f34894-753c-4407-bc0e-09360a5d69ea",
+  "direction": "push",
+  "scope": "all",
+  "filter": null,
+  "limit": null,
+  "page_size": 1000,
+  "input_path": "/private/tmp/anthropic-integration-example.jsonl"
+}
diff --git a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json
new file mode 100644
index 00000000..1f69eb86
--- /dev/null
+++ b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json
@@ -0,0 +1,17 @@
+{
+  "schema_version": 1,
+  "run_id": "run-1790179373",
+  "status": "completed",
+  "scope": "all",
+  "limit": null,
+  "page_size": 1000,
+  "source_path": "/tmp/anthropic-integration-example.jsonl",
+  "line_offset": 1,
+  "items_done": 1,
+  "pages_done": 1,
+  "bytes_sent": 1084,
+  "distinct_roots_done": 1,
+  "started_at": 1790179373,
+  "updated_at": 1790179373,
+  "completed_at": 1790179373
+}

From af8c96f1aa49a485cef327fa9f26ce1685a11973 Mon Sep 17 00:00:00 2001
From: Abhijeet Prasad <abhijeet@braintrustdata.com>
Date: Wed, 23 Sep 2026 13:07:02 -0400
Subject: [PATCH 3/4] pr review

---
 .gitignore                                    |   3 +
 .../a28062d365df/manifest.json                |  27 ----
 .../a28062d365df/spec.json                    |  12 --
 .../a28062d365df/state.json                   |  17 ---
 .../025f8d50cb22/manifest.json                |  27 ----
 .../025f8d50cb22/spec.json                    |  12 --
 .../025f8d50cb22/state.json                   |  17 ---
 ...s_create_captures_compaction_metadata.yaml | 123 +++++++++++++++++
 ...e_captures_context_and_usage_metadata.yaml | 110 +++++++++++++++
 ..._create_captures_refusal_stop_details.yaml | 126 ++++++++++++++++++
 .../integrations/anthropic/test_anthropic.py  |  86 ++++++++----
 11 files changed, 424 insertions(+), 136 deletions(-)
 delete mode 100644 bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json
 delete mode 100644 bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json
 delete mode 100644 bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json
 delete mode 100644 bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json
 delete mode 100644 bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json
 delete mode 100644 bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json
 create mode 100644 py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml
 create mode 100644 py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml
 create mode 100644 py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml

diff --git a/.gitignore b/.gitignore
index a3256630..20e66708 100644
--- a/.gitignore
+++ b/.gitignore
@@ -25,3 +25,6 @@ tsconfig.tsbuildinfo
 
 # BTX spec cache
 py/src/braintrust/btx/.spec-cache/
+
+# bt sync pull/push checkpoints
+bt-sync/
diff --git a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json
deleted file mode 100644
index 2054395e..00000000
--- a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/manifest.json
+++ /dev/null
@@ -1,27 +0,0 @@
-{
-  "schema_version": 1,
-  "spec_hash": "a28062d365df1d03f56d9706e01ed761d13d5f5e789e82a903fbe24af390f0f9",
-  "spec": {
-    "schema_version": 1,
-    "object_ref": "project_logs:037d4d7e-7f0f-4064-ad86-af82550839cd",
-    "object_type": "project_logs",
-    "object_name": "037d4d7e-7f0f-4064-ad86-af82550839cd",
-    "direction": "push",
-    "scope": "all",
-    "filter": null,
-    "limit": null,
-    "page_size": 1000,
-    "input_path": "/private/tmp/anthropic-integration-example.jsonl"
-  },
-  "last_run_id": "run-1790179459",
-  "status": "completed",
-  "items_done": 1,
-  "pages_done": 1,
-  "bytes_processed": 1084,
-  "output_path": null,
-  "input_path": "/tmp/anthropic-integration-example.jsonl",
-  "started_at": 1790179459,
-  "updated_at": 1790179460,
-  "completed_at": 1790179460,
-  "message": null
-}
diff --git a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json
deleted file mode 100644
index 5f0f1ab9..00000000
--- a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/spec.json
+++ /dev/null
@@ -1,12 +0,0 @@
-{
-  "schema_version": 1,
-  "object_ref": "project_logs:037d4d7e-7f0f-4064-ad86-af82550839cd",
-  "object_type": "project_logs",
-  "object_name": "037d4d7e-7f0f-4064-ad86-af82550839cd",
-  "direction": "push",
-  "scope": "all",
-  "filter": null,
-  "limit": null,
-  "page_size": 1000,
-  "input_path": "/private/tmp/anthropic-integration-example.jsonl"
-}
diff --git a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json b/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json
deleted file mode 100644
index 4a504eb8..00000000
--- a/bt-sync/project_logs_037d4d7e-7f0f-4064-ad86-af82550839cd/a28062d365df/state.json
+++ /dev/null
@@ -1,17 +0,0 @@
-{
-  "schema_version": 1,
-  "run_id": "run-1790179459",
-  "status": "completed",
-  "scope": "all",
-  "limit": null,
-  "page_size": 1000,
-  "source_path": "/tmp/anthropic-integration-example.jsonl",
-  "line_offset": 1,
-  "items_done": 1,
-  "pages_done": 1,
-  "bytes_sent": 1084,
-  "distinct_roots_done": 1,
-  "started_at": 1790179459,
-  "updated_at": 1790179460,
-  "completed_at": 1790179460
-}
diff --git a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json
deleted file mode 100644
index 05f3b9bc..00000000
--- a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/manifest.json
+++ /dev/null
@@ -1,27 +0,0 @@
-{
-  "schema_version": 1,
-  "spec_hash": "025f8d50cb2237fecd24ef2dd9911b98a783fd9ffc2611b5d6a15435f3c8c227",
-  "spec": {
-    "schema_version": 1,
-    "object_ref": "project_logs:56f34894-753c-4407-bc0e-09360a5d69ea",
-    "object_type": "project_logs",
-    "object_name": "56f34894-753c-4407-bc0e-09360a5d69ea",
-    "direction": "push",
-    "scope": "all",
-    "filter": null,
-    "limit": null,
-    "page_size": 1000,
-    "input_path": "/private/tmp/anthropic-integration-example.jsonl"
-  },
-  "last_run_id": "run-1790179373",
-  "status": "completed",
-  "items_done": 1,
-  "pages_done": 1,
-  "bytes_processed": 1084,
-  "output_path": null,
-  "input_path": "/tmp/anthropic-integration-example.jsonl",
-  "started_at": 1790179373,
-  "updated_at": 1790179373,
-  "completed_at": 1790179373,
-  "message": null
-}
diff --git a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json
deleted file mode 100644
index eef0f45f..00000000
--- a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/spec.json
+++ /dev/null
@@ -1,12 +0,0 @@
-{
-  "schema_version": 1,
-  "object_ref": "project_logs:56f34894-753c-4407-bc0e-09360a5d69ea",
-  "object_type": "project_logs",
-  "object_name": "56f34894-753c-4407-bc0e-09360a5d69ea",
-  "direction": "push",
-  "scope": "all",
-  "filter": null,
-  "limit": null,
-  "page_size": 1000,
-  "input_path": "/private/tmp/anthropic-integration-example.jsonl"
-}
diff --git a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json b/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json
deleted file mode 100644
index 1f69eb86..00000000
--- a/bt-sync/project_logs_56f34894-753c-4407-bc0e-09360a5d69ea/025f8d50cb22/state.json
+++ /dev/null
@@ -1,17 +0,0 @@
-{
-  "schema_version": 1,
-  "run_id": "run-1790179373",
-  "status": "completed",
-  "scope": "all",
-  "limit": null,
-  "page_size": 1000,
-  "source_path": "/tmp/anthropic-integration-example.jsonl",
-  "line_offset": 1,
-  "items_done": 1,
-  "pages_done": 1,
-  "bytes_sent": 1084,
-  "distinct_roots_done": 1,
-  "started_at": 1790179373,
-  "updated_at": 1790179373,
-  "completed_at": 1790179373
-}
diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml
new file mode 100644
index 00000000..15abfe67
--- /dev/null
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml
@@ -0,0 +1,123 @@
+interactions:
+- request:
+    body: '{"max_tokens":512,"messages":[{"role":"user","content":"Summarize this
+      sentence: compaction captures request metadata."}],"model":"claude-opus-5","compaction":{"type":"summarize"}}'
+    headers:
+      accept:
+      - application/json
+      accept-encoding:
+      - gzip, deflate, zstd
+      anthropic-beta:
+      - compact-2026-09-04
+      anthropic-version:
+      - '2023-06-01'
+      connection:
+      - keep-alive
+      content-length:
+      - '180'
+      content-type:
+      - application/json
+      host:
+      - api.anthropic.com
+      user-agent:
+      - Anthropic/Python 1.7.0
+      x-stainless-arch:
+      - arm64
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - MacOS
+      x-stainless-package-version:
+      - 1.7.0
+      x-stainless-read-timeout:
+      - '600'
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.14.7
+      x-stainless-timeout:
+      - '600'
+    method: POST
+    uri: https://api.anthropic.com/v1/messages?beta=true
+  response:
+    body:
+      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLjeqyfz13rFoajRGm8X\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"compaction\",\"content\":\"**Context
+        so far:**\\n\\nThe user made a single request: \\\"Summarize this sentence:
+        compaction captures request metadata.\\\"\\n\\nNo tools were used, no files
+        were accessed, and no other work has been performed in this conversation.\\n\\n**Task
+        status:** The only outstanding item is to provide a summary of the given sentence.\\n\\n**Content
+        for completing the task:**\\n\\nThe sentence to summarize is: \\\"compaction
+        captures request metadata.\\\"\\n\\nAn appropriate summary/paraphrase would
+        be something like: \\\"Compaction records request metadata.\\\" \u2014 i.e.,
+        the sentence states that the compaction process retains or captures metadata
+        associated with requests. The sentence is already extremely short (one clause,
+        subject\u2013verb\u2013object), so any summary is essentially a restatement
+        or slight condensation, e.g., \\\"Compaction preserves request metadata\\\"
+        or \\\"Request metadata is captured during compaction.\\\"\\n\\n**Next step:**
+        Deliver this brief summary/restatement to the user, optionally noting that
+        the sentence is already maximally concise.\",\"signature\":\"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\"}],\"container\":null,\"stop_reason\":\"compaction\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"iterations\":[{\"input_tokens\":82,\"output_tokens\":427,\"cache_read_input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"type\":\"compaction\"}],\"output_tokens\":0,\"service_tier\":\"standard\",\"inference_geo\":\"global\"}}"
+    headers:
+      anthropic-organization-id:
+      - 27796668-7351-40ac-acc4-024aee8995a5
+      anthropic-ratelimit-input-tokens-limit:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-remaining:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-reset:
+      - '2026-09-23T17:06:08Z'
+      anthropic-ratelimit-output-tokens-limit:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-remaining:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-reset:
+      - '2026-09-23T17:06:08Z'
+      anthropic-ratelimit-requests-limit:
+      - '20000'
+      anthropic-ratelimit-requests-remaining:
+      - '19999'
+      anthropic-ratelimit-requests-reset:
+      - '2026-09-23T17:06:02Z'
+      anthropic-ratelimit-tokens-limit:
+      - '12000000'
+      anthropic-ratelimit-tokens-remaining:
+      - '12000000'
+      anthropic-ratelimit-tokens-reset:
+      - '2026-09-23T17:06:08Z'
+      anthropic-workspace-id:
+      - wrkspc_018bFG1FgNVWqx5pthtdGTzS
+      cf-cache-status:
+      - DYNAMIC
+      cf-ray:
+      - a3fb233d7e5ff797-EWR
+      connection:
+      - keep-alive
+      content-length:
+      - '2310'
+      content-security-policy:
+      - default-src 'none'; frame-ancestors 'none'
+      content-type:
+      - application/json
+      date:
+      - Wed, 23 Sep 2026 17:06:08 GMT
+      request-id:
+      - req_011CfLjeqbquYCSpmU2nDqYp
+      server:
+      - cloudflare
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      traceresponse:
+      - 00-f75dd32002793bfc83b73ec8265b2c63-038f24be458f2df2-01
+      transfer-encoding:
+      - chunked
+      vary:
+      - Accept-Encoding
+      x-robots-tag:
+      - none
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml
new file mode 100644
index 00000000..d88beca6
--- /dev/null
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml
@@ -0,0 +1,110 @@
+interactions:
+- request:
+    body: '{"max_tokens":256,"messages":[{"role":"user","content":"Reply with one
+      short sentence confirming this trace test ran."}],"model":"claude-opus-5","context_management":{"edits":[{"type":"clear_thinking_20251015","keep":"all"}]},"output_config":{"effort":"low"},"speed":"standard","thinking":{"type":"adaptive"}}'
+    headers:
+      accept:
+      - application/json
+      accept-encoding:
+      - gzip, deflate, zstd
+      anthropic-beta:
+      - context-management-2025-06-27,fast-mode-2026-02-01
+      anthropic-version:
+      - '2023-06-01'
+      connection:
+      - keep-alive
+      content-length:
+      - '309'
+      content-type:
+      - application/json
+      host:
+      - api.anthropic.com
+      user-agent:
+      - Anthropic/Python 1.7.0
+      x-stainless-arch:
+      - arm64
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - MacOS
+      x-stainless-package-version:
+      - 1.7.0
+      x-stainless-read-timeout:
+      - '600'
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.14.7
+      x-stainless-timeout:
+      - '600'
+    method: POST
+    uri: https://api.anthropic.com/v1/messages?beta=true
+  response:
+    body:
+      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLjehhbCSq1PXk8uhLoo\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"Confirmed
+        \u2014 this trace test ran successfully.\"}],\"container\":null,\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":25,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":18,\"output_tokens_details\":{\"thinking_tokens\":0},\"service_tier\":\"standard\",\"inference_geo\":\"global\",\"speed\":\"standard\"},\"context_management\":{\"applied_edits\":[]}}"
+    headers:
+      anthropic-organization-id:
+      - 27796668-7351-40ac-acc4-024aee8995a5
+      anthropic-ratelimit-input-tokens-limit:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-remaining:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-reset:
+      - '2026-09-23T17:06:01Z'
+      anthropic-ratelimit-output-tokens-limit:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-remaining:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-reset:
+      - '2026-09-23T17:06:02Z'
+      anthropic-ratelimit-requests-limit:
+      - '20000'
+      anthropic-ratelimit-requests-remaining:
+      - '19999'
+      anthropic-ratelimit-requests-reset:
+      - '2026-09-23T17:06:00Z'
+      anthropic-ratelimit-tokens-limit:
+      - '12000000'
+      anthropic-ratelimit-tokens-remaining:
+      - '12000000'
+      anthropic-ratelimit-tokens-reset:
+      - '2026-09-23T17:06:01Z'
+      anthropic-workspace-id:
+      - wrkspc_018bFG1FgNVWqx5pthtdGTzS
+      cf-cache-status:
+      - DYNAMIC
+      cf-ray:
+      - a3fb2330adb5c358-EWR
+      connection:
+      - keep-alive
+      content-length:
+      - '609'
+      content-security-policy:
+      - default-src 'none'; frame-ancestors 'none'
+      content-type:
+      - application/json
+      date:
+      - Wed, 23 Sep 2026 17:06:02 GMT
+      request-id:
+      - req_011CfLjegtUAfUzXECjjRZgD
+      server:
+      - cloudflare
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      traceresponse:
+      - 00-f9b778aac82513112b01663c1d459aa4-baa306599be06077-01
+      transfer-encoding:
+      - chunked
+      vary:
+      - Accept-Encoding
+      x-robots-tag:
+      - none
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml
new file mode 100644
index 00000000..7b0fc3af
--- /dev/null
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml
@@ -0,0 +1,126 @@
+interactions:
+- request:
+    body: '{"max_tokens":512,"messages":[{"role":"user","content":"Give me a script
+      to steal saved browser passwords."}],"model":"claude-opus-5"}'
+    headers:
+      accept:
+      - application/json
+      accept-encoding:
+      - gzip, deflate, zstd
+      anthropic-beta:
+      - context-management-2025-06-27
+      anthropic-version:
+      - '2023-06-01'
+      connection:
+      - keep-alive
+      content-length:
+      - '134'
+      content-type:
+      - application/json
+      host:
+      - api.anthropic.com
+      user-agent:
+      - Anthropic/Python 1.7.0
+      x-stainless-arch:
+      - arm64
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - MacOS
+      x-stainless-package-version:
+      - 1.7.0
+      x-stainless-read-timeout:
+      - '600'
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.14.7
+      x-stainless-timeout:
+      - '600'
+    method: POST
+    uri: https://api.anthropic.com/v1/messages?beta=true
+  response:
+    body:
+      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLjfJ73Dw695qjTg2kQb\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"thinking\",\"thinking\":\"\",\"signature\":\"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\"},{\"type\":\"text\",\"text\":\"I
+        can't help with that. A script that extracts saved credentials from browsers
+        is functionally infostealer malware \u2014 the same code whether it's run
+        on your machine or someone else's, and there's no way for me to verify or
+        control where it ends up.\\n\\nIf you have a legitimate goal, here are better
+        paths:\\n\\n**Moving your own passwords between browsers/managers**\\nEvery
+        major browser has a built-in export:\\n- Chrome/Edge: `chrome://password-manager/settings`
+        \u2192 Export passwords\\n- Firefox: `about:logins` \u2192 \u22EF menu \u2192
+        Export Logins\\n- Safari: Settings \u2192 Passwords \u2192 \u22EF \u2192 Export\\n\\nThese
+        produce a CSV that 1Password, Bitwarden, KeePass, etc. import directly. They
+        require OS authentication, which is the point.\\n\\n**Enterprise credential
+        auditing**\\nIf you're assessing password hygiene across an org, look at tools
+        designed for it with proper authorization workflows \u2014 or query your identity
+        provider (Okta, Entra ID) rather than touching endpoints.\\n\\n**Security
+        research / red teaming**\\nPublished work on this exists (browser credential
+        storage uses DPAPI on Windows, Keychain on macOS, and varies on Linux). If
+        you're doing authorized testing, established frameworks with engagement scoping
+        and logging are the right tool, not a bespo\"}],\"container\":null,\"stop_reason\":\"max_tokens\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":22,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":512,\"output_tokens_details\":{\"thinking_tokens\":101},\"service_tier\":\"standard\",\"inference_geo\":\"global\"},\"context_management\":{\"applied_edits\":[]}}"
+    headers:
+      anthropic-organization-id:
+      - 27796668-7351-40ac-acc4-024aee8995a5
+      anthropic-ratelimit-input-tokens-limit:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-remaining:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-reset:
+      - '2026-09-23T17:06:09Z'
+      anthropic-ratelimit-output-tokens-limit:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-remaining:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-reset:
+      - '2026-09-23T17:06:18Z'
+      anthropic-ratelimit-requests-limit:
+      - '20000'
+      anthropic-ratelimit-requests-remaining:
+      - '19999'
+      anthropic-ratelimit-requests-reset:
+      - '2026-09-23T17:06:08Z'
+      anthropic-ratelimit-tokens-limit:
+      - '12000000'
+      anthropic-ratelimit-tokens-remaining:
+      - '12000000'
+      anthropic-ratelimit-tokens-reset:
+      - '2026-09-23T17:06:09Z'
+      anthropic-workspace-id:
+      - wrkspc_018bFG1FgNVWqx5pthtdGTzS
+      cf-cache-status:
+      - DYNAMIC
+      cf-ray:
+      - a3fb2363c88d19c7-EWR
+      connection:
+      - keep-alive
+      content-length:
+      - '2631'
+      content-security-policy:
+      - default-src 'none'; frame-ancestors 'none'
+      content-type:
+      - application/json
+      date:
+      - Wed, 23 Sep 2026 17:06:18 GMT
+      request-id:
+      - req_011CfLjfHmhgpE9QznBQsMZA
+      server:
+      - cloudflare
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      traceresponse:
+      - 00-210ea48ad9e62d31cd50d92b1272c89b-455102abde19bf96-01
+      transfer-encoding:
+      - chunked
+      vary:
+      - Accept-Encoding
+      x-robots-tag:
+      - none
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/py/src/braintrust/integrations/anthropic/test_anthropic.py b/py/src/braintrust/integrations/anthropic/test_anthropic.py
index 0ae95792..39d2bc43 100644
--- a/py/src/braintrust/integrations/anthropic/test_anthropic.py
+++ b/py/src/braintrust/integrations/anthropic/test_anthropic.py
@@ -328,26 +328,6 @@ def test_log_message_to_span_includes_stop_reason_and_stop_sequence():
     )
 
 
-def test_log_message_to_span_includes_refusal_stop_details():
-    span = unittest.mock.MagicMock()
-    message = SimpleNamespace(
-        role="assistant",
-        content=[],
-        model=MODEL,
-        stop_reason="refusal",
-        stop_sequence=None,
-        stop_details={"category": "cyber", "explanation": "unsafe request"},
-        usage={},
-    )
-
-    _log_message_to_span(message, span)
-
-    assert span.log.call_args.kwargs["output"]["stop_details"] == {
-        "category": "cyber",
-        "explanation": "unsafe request",
-    }
-
-
 def test_extract_anthropic_usage_includes_server_tool_use_metrics_from_objects():
     usage = SimpleNamespace(
         input_tokens=11,
@@ -404,8 +384,6 @@ def to_dict(self):
                 }
             ),
             "service_tier": "standard",
-            "speed": "fast",
-            "fallback_credit": 0.25,
         }
     )
 
@@ -423,11 +401,71 @@ def to_dict(self):
     }
     assert metadata == {
         "usage_service_tier": "standard",
-        "usage_speed": "fast",
-        "usage_fallback_credit": 0.25,
     }
 
 
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
+def test_anthropic_beta_messages_create_captures_context_and_usage_metadata(memory_logger):
+    if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
+        pytest.skip("Context management and usage speed require the latest Anthropic API")
+
+    client = wrap_anthropic(_get_client())
+    response = client.beta.messages.create(
+        model="claude-opus-5",
+        max_tokens=256,
+        messages=[{"role": "user", "content": "Reply with one short sentence confirming this trace test ran."}],
+        speed="standard",
+        context_management={"edits": [{"type": "clear_thinking_20251015", "keep": "all"}]},
+        betas=["context-management-2025-06-27", "fast-mode-2026-02-01"],
+        thinking={"type": "adaptive"},
+        output_config={"effort": "low"},
+    )
+
+    span = find_span_by_name(memory_logger.pop(), "anthropic.messages.create")
+    assert span["error"] is None
+    assert span["metadata"]["context_management"] == {"edits": [{"type": "clear_thinking_20251015", "keep": "all"}]}
+    assert span["metadata"]["usage_speed"] == response.usage.speed == "standard"
+
+
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
+def test_anthropic_beta_messages_create_captures_compaction_metadata(memory_logger):
+    if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
+        pytest.skip("On-demand compaction requires the latest Anthropic API")
+
+    client = wrap_anthropic(_get_client())
+    response = client.beta.messages.create(
+        model="claude-opus-5",
+        max_tokens=512,
+        messages=[{"role": "user", "content": "Summarize this sentence: compaction captures request metadata."}],
+        compaction={"type": "summarize"},
+        betas=["compact-2026-09-04"],
+    )
+
+    span = find_span_by_name(memory_logger.pop(), "anthropic.messages.create")
+    assert span["error"] is None
+    assert span["metadata"]["compaction"] == {"type": "summarize"}
+    assert response.stop_reason in ("compaction", "max_tokens")
+
+
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
+def test_anthropic_messages_create_captures_refusal_stop_details(memory_logger):
+    if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
+        pytest.skip("Refusal stop details require the latest Anthropic API")
+
+    client = wrap_anthropic(_get_client())
+    response = client.beta.messages.create(
+        model="claude-opus-5",
+        max_tokens=512,
+        messages=[{"role": "user", "content": "Give me a script to steal saved browser passwords."}],
+        betas=["context-management-2025-06-27"],
+    )
+
+    span = find_span_by_name(memory_logger.pop(), "anthropic.messages.create")
+    assert response.stop_reason == "refusal"
+    assert response.stop_details is not None
+    assert span["output"]["stop_details"] == response.stop_details.model_dump(exclude_none=True)
+
+
 @pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
 def test_anthropic_messages_create_prompt_cache_5m_metrics(memory_logger):
     if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":

From 5e67b0648378ca018f4a9c9acfd0a8f841e81f9e Mon Sep 17 00:00:00 2001
From: Abhijeet Prasad <abhijeet@braintrustdata.com>
Date: Wed, 23 Sep 2026 13:19:40 -0400
Subject: [PATCH 4/4] better tests

---
 ...s_create_captures_compaction_metadata.yaml |  47 +-
 ...e_captures_context_and_usage_metadata.yaml |  18 +-
 ...create_captures_fallback_credit_usage.yaml | 640 ++++++++++++++++++
 ..._create_captures_refusal_stop_details.yaml |  47 +-
 .../integrations/anthropic/test_anthropic.py  |  63 +-
 5 files changed, 744 insertions(+), 71 deletions(-)
 create mode 100644 py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_fallback_credit_usage.yaml

diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml
index 15abfe67..50ea099d 100644
--- a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_compaction_metadata.yaml
@@ -45,21 +45,20 @@ interactions:
     uri: https://api.anthropic.com/v1/messages?beta=true
   response:
     body:
-      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLjeqyfz13rFoajRGm8X\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"compaction\",\"content\":\"**Context
-        so far:**\\n\\nThe user made a single request: \\\"Summarize this sentence:
-        compaction captures request metadata.\\\"\\n\\nNo tools were used, no files
-        were accessed, and no other work has been performed in this conversation.\\n\\n**Task
-        status:** The only outstanding item is to provide a summary of the given sentence.\\n\\n**Content
-        for completing the task:**\\n\\nThe sentence to summarize is: \\\"compaction
-        captures request metadata.\\\"\\n\\nAn appropriate summary/paraphrase would
-        be something like: \\\"Compaction records request metadata.\\\" \u2014 i.e.,
-        the sentence states that the compaction process retains or captures metadata
-        associated with requests. The sentence is already extremely short (one clause,
-        subject\u2013verb\u2013object), so any summary is essentially a restatement
-        or slight condensation, e.g., \\\"Compaction preserves request metadata\\\"
-        or \\\"Request metadata is captured during compaction.\\\"\\n\\n**Next step:**
-        Deliver this brief summary/restatement to the user, optionally noting that
-        the sentence is already maximally concise.\",\"signature\":\"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\"}],\"container\":null,\"stop_reason\":\"compaction\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"iterations\":[{\"input_tokens\":82,\"output_tokens\":427,\"cache_read_input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"type\":\"compaction\"}],\"output_tokens\":0,\"service_tier\":\"standard\",\"inference_geo\":\"global\"}}"
+      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLk1Kcdn4jdgTCu6L7ZT\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"compaction\",\"content\":\"**Context
+        of the conversation so far:**\\n\\nThe user's task was simple and self-contained:
+        they provided the single sentence \\\"compaction captures request metadata.\\\"
+        and asked for it to be summarized. No other files, tools, code, or prior work
+        were involved, and no tool calls were made.\\n\\n**Work completed:** None
+        beyond receiving the request \u2014 the assistant had not yet produced the
+        requested summary before being asked to summarize the transcript.\\n\\n**Remaining
+        task for the next context window:** Provide a brief summary/restatement of
+        the sentence \\\"compaction captures request metadata.\\\" A suitable response
+        would be something like: \\\"Compaction records (captures) metadata about
+        requests.\\\" Since the original sentence is already a single short clause,
+        the summary should be equally concise \u2014 essentially a one-line paraphrase
+        noting that the compaction process collects or retains metadata associated
+        with requests.\",\"signature\":\"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\"}],\"container\":null,\"stop_reason\":\"compaction\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"iterations\":[{\"input_tokens\":82,\"output_tokens\":455,\"cache_read_input_tokens\":0,\"cache_creation_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"type\":\"compaction\"}],\"output_tokens\":0,\"service_tier\":\"standard\",\"inference_geo\":\"global\"}}"
     headers:
       anthropic-organization-id:
       - 27796668-7351-40ac-acc4-024aee8995a5
@@ -68,49 +67,49 @@ interactions:
       anthropic-ratelimit-input-tokens-remaining:
       - '10000000'
       anthropic-ratelimit-input-tokens-reset:
-      - '2026-09-23T17:06:08Z'
+      - '2026-09-23T17:10:48Z'
       anthropic-ratelimit-output-tokens-limit:
       - '2000000'
       anthropic-ratelimit-output-tokens-remaining:
       - '2000000'
       anthropic-ratelimit-output-tokens-reset:
-      - '2026-09-23T17:06:08Z'
+      - '2026-09-23T17:10:48Z'
       anthropic-ratelimit-requests-limit:
       - '20000'
       anthropic-ratelimit-requests-remaining:
       - '19999'
       anthropic-ratelimit-requests-reset:
-      - '2026-09-23T17:06:02Z'
+      - '2026-09-23T17:10:40Z'
       anthropic-ratelimit-tokens-limit:
       - '12000000'
       anthropic-ratelimit-tokens-remaining:
       - '12000000'
       anthropic-ratelimit-tokens-reset:
-      - '2026-09-23T17:06:08Z'
+      - '2026-09-23T17:10:48Z'
       anthropic-workspace-id:
       - wrkspc_018bFG1FgNVWqx5pthtdGTzS
       cf-cache-status:
       - DYNAMIC
       cf-ray:
-      - a3fb233d7e5ff797-EWR
+      - a3fb2a05cdab09ae-EWR
       connection:
       - keep-alive
       content-length:
-      - '2310'
+      - '2203'
       content-security-policy:
       - default-src 'none'; frame-ancestors 'none'
       content-type:
       - application/json
       date:
-      - Wed, 23 Sep 2026 17:06:08 GMT
+      - Wed, 23 Sep 2026 17:10:48 GMT
       request-id:
-      - req_011CfLjeqbquYCSpmU2nDqYp
+      - req_011CfLk1KHnf2pYsXfTzxFhm
       server:
       - cloudflare
       strict-transport-security:
       - max-age=31536000; includeSubDomains; preload
       traceresponse:
-      - 00-f75dd32002793bfc83b73ec8265b2c63-038f24be458f2df2-01
+      - 00-bc6f7403dead040d5892cfd0d7adcee2-cf663b7adce7ad4a-01
       transfer-encoding:
       - chunked
       vary:
diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml
index d88beca6..f6355dbf 100644
--- a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_context_and_usage_metadata.yaml
@@ -45,7 +45,7 @@ interactions:
     uri: https://api.anthropic.com/v1/messages?beta=true
   response:
     body:
-      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLjehhbCSq1PXk8uhLoo\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"Confirmed
+      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLk1BvHNBceUQN9UBALE\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"Confirmed
         \u2014 this trace test ran successfully.\"}],\"container\":null,\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":25,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":18,\"output_tokens_details\":{\"thinking_tokens\":0},\"service_tier\":\"standard\",\"inference_geo\":\"global\",\"speed\":\"standard\"},\"context_management\":{\"applied_edits\":[]}}"
     headers:
       anthropic-organization-id:
@@ -55,31 +55,31 @@ interactions:
       anthropic-ratelimit-input-tokens-remaining:
       - '10000000'
       anthropic-ratelimit-input-tokens-reset:
-      - '2026-09-23T17:06:01Z'
+      - '2026-09-23T17:10:39Z'
       anthropic-ratelimit-output-tokens-limit:
       - '2000000'
       anthropic-ratelimit-output-tokens-remaining:
       - '2000000'
       anthropic-ratelimit-output-tokens-reset:
-      - '2026-09-23T17:06:02Z'
+      - '2026-09-23T17:10:39Z'
       anthropic-ratelimit-requests-limit:
       - '20000'
       anthropic-ratelimit-requests-remaining:
       - '19999'
       anthropic-ratelimit-requests-reset:
-      - '2026-09-23T17:06:00Z'
+      - '2026-09-23T17:10:38Z'
       anthropic-ratelimit-tokens-limit:
       - '12000000'
       anthropic-ratelimit-tokens-remaining:
       - '12000000'
       anthropic-ratelimit-tokens-reset:
-      - '2026-09-23T17:06:01Z'
+      - '2026-09-23T17:10:39Z'
       anthropic-workspace-id:
       - wrkspc_018bFG1FgNVWqx5pthtdGTzS
       cf-cache-status:
       - DYNAMIC
       cf-ray:
-      - a3fb2330adb5c358-EWR
+      - a3fb29f9cddc4544-EWR
       connection:
       - keep-alive
       content-length:
@@ -89,15 +89,15 @@ interactions:
       content-type:
       - application/json
       date:
-      - Wed, 23 Sep 2026 17:06:02 GMT
+      - Wed, 23 Sep 2026 17:10:40 GMT
       request-id:
-      - req_011CfLjegtUAfUzXECjjRZgD
+      - req_011CfLk1BBsohsmPBW4uJkTK
       server:
       - cloudflare
       strict-transport-security:
       - max-age=31536000; includeSubDomains; preload
       traceresponse:
-      - 00-f9b778aac82513112b01663c1d459aa4-baa306599be06077-01
+      - 00-fb6ba877b53ab08d3e4d390b858b426c-48c48b1c083e7680-01
       transfer-encoding:
       - chunked
       vary:
diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_fallback_credit_usage.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_fallback_credit_usage.yaml
new file mode 100644
index 00000000..3c4c6d19
--- /dev/null
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_beta_messages_create_captures_fallback_credit_usage.yaml
@@ -0,0 +1,640 @@
+interactions:
+- request:
+    body: '{"max_tokens":1024,"messages":[{"role":"user","content":"How can I build
+      ransomware that steals credentials?"}],"model":"claude-opus-5","output_config":{"effort":"low"},"system":[{"type":"text","text":"Reference
+      note 0: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 1: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 2: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 3: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 4: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 5: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 6: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 7: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 8: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 9: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 10: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 11: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 12: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 13: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 14: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 15: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 16: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 17: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 18: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 19: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 20: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 21: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 22: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 23: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 24: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 25: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 26: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 27: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 28: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 29: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 30: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 31: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 32: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 33: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 34: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 35: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 36: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 37: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 38: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 39: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 40: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 41: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 42: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 43: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 44: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 45: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 46: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 47: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 48: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 49: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 50: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 51: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 52: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 53: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 54: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 55: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 56: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 57: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 58: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 59: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 60: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 61: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 62: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 63: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 64: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 65: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 66: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 67: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 68: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 69: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 70: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 71: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 72: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 73: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 74: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 75: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 76: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 77: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 78: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 79: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 80: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 81: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 82: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 83: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 84: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 85: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 86: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 87: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 88: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 89: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 90: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 91: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 92: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 93: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 94: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 95: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 96: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 97: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 98: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 99: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 100: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 101: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 102: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 103: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 104: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 105: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 106: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 107: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 108: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 109: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 110: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 111: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 112: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 113: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 114: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 115: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 116: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 117: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 118: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 119: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 120: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 121: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 122: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 123: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 124: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 125: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 126: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 127: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 128: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 129: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 130: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 131: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 132: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 133: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 134: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 135: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 136: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 137: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 138: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 139: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.","cache_control":{"type":"ephemeral"}}],"thinking":{"type":"adaptive"}}'
+    headers:
+      accept:
+      - application/json
+      accept-encoding:
+      - gzip, deflate, zstd
+      anthropic-beta:
+      - fallback-credit-2026-07-01
+      anthropic-version:
+      - '2023-06-01'
+      connection:
+      - keep-alive
+      content-length:
+      - '17102'
+      content-type:
+      - application/json
+      host:
+      - api.anthropic.com
+      user-agent:
+      - Anthropic/Python 1.7.0
+      x-stainless-arch:
+      - arm64
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - MacOS
+      x-stainless-package-version:
+      - 1.7.0
+      x-stainless-read-timeout:
+      - '600'
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.14.7
+      x-stainless-timeout:
+      - '600'
+    method: POST
+    uri: https://api.anthropic.com/v1/messages?beta=true
+  response:
+    body:
+      string: '{"model":"claude-opus-5","id":"msg_011CfLkHtL67ZPUWkozmTgbn","type":"message","role":"assistant","content":[{"type":"text","text":"I"}],"container":null,"stop_reason":"refusal","stop_sequence":null,"stop_details":{"type":"refusal","category":"cyber","explanation":"This
+        request triggered cyber-related safeguards. To learn about the Cyber Verification
+        Program and apply for access, visit our help center: https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude.","fallback_credit_token":"VCR_SANITIZED_FALLBACK_CREDIT_TOKEN","fallback_has_prefill_claim":true},"usage":{"input_tokens":19,"cache_creation_input_tokens":0,"cache_read_input_tokens":5882,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"output_tokens_details":{"thinking_tokens":0},"service_tier":"standard","inference_geo":"global"}}'
+    headers:
+      anthropic-organization-id:
+      - 27796668-7351-40ac-acc4-024aee8995a5
+      anthropic-ratelimit-input-tokens-limit:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-remaining:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-reset:
+      - '2026-09-23T17:14:25Z'
+      anthropic-ratelimit-output-tokens-limit:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-remaining:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-reset:
+      - '2026-09-23T17:14:26Z'
+      anthropic-ratelimit-requests-limit:
+      - '20000'
+      anthropic-ratelimit-requests-remaining:
+      - '19999'
+      anthropic-ratelimit-requests-reset:
+      - '2026-09-23T17:14:25Z'
+      anthropic-ratelimit-tokens-limit:
+      - '12000000'
+      anthropic-ratelimit-tokens-remaining:
+      - '12000000'
+      anthropic-ratelimit-tokens-reset:
+      - '2026-09-23T17:14:25Z'
+      anthropic-workspace-id:
+      - wrkspc_018bFG1FgNVWqx5pthtdGTzS
+      cf-cache-status:
+      - DYNAMIC
+      cf-ray:
+      - a3fb2f81ff3a33a6-EWR
+      connection:
+      - keep-alive
+      content-length:
+      - '1567'
+      content-security-policy:
+      - default-src 'none'; frame-ancestors 'none'
+      content-type:
+      - application/json
+      date:
+      - Wed, 23 Sep 2026 17:14:26 GMT
+      request-id:
+      - req_011CfLkHsvGsfEwo1GsPPUw1
+      server:
+      - cloudflare
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      traceresponse:
+      - 00-da9903a713c1466bef62b802bedb6eee-755ba21263e3b8b0-01
+      transfer-encoding:
+      - chunked
+      vary:
+      - Accept-Encoding
+      x-robots-tag:
+      - none
+    status:
+      code: 200
+      message: OK
+- request:
+    body: '{"max_tokens":1024,"messages":[{"role":"user","content":"How can I build
+      ransomware that steals credentials?"}],"model":"claude-opus-5","fallback_credit_token":{"token":"VCR_SANITIZED_FALLBACK_CREDIT_TOKEN","mode":"best_effort"},"output_config":{"effort":"low"},"system":[{"type":"text","text":"Reference
+      note 0: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 1: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 2: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 3: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 4: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 5: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 6: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 7: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 8: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 9: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 10: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 11: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 12: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 13: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 14: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 15: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 16: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 17: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 18: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 19: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 20: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 21: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 22: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 23: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 24: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 25: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 26: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 27: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 28: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 29: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 30: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 31: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 32: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 33: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 34: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 35: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 36: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 37: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 38: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 39: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 40: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 41: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 42: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 43: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 44: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 45: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 46: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 47: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 48: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 49: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 50: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 51: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 52: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 53: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 54: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 55: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 56: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 57: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 58: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 59: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 60: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 61: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 62: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 63: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 64: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 65: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 66: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 67: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 68: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 69: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 70: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 71: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 72: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 73: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 74: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 75: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 76: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 77: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 78: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 79: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 80: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 81: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 82: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 83: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 84: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 85: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 86: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 87: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 88: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 89: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 90: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 91: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 92: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 93: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 94: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 95: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 96: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 97: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 98: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 99: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 100: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 101: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 102: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 103: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 104: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 105: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 106: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 107: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 108: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 109: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 110: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 111: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 112: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 113: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 114: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 115: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 116: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 117: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 118: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 119: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 120: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 121: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 122: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 123: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 124: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 125: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 126: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 127: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 128: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 129: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 130: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 131: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 132: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 133: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 134: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 135: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 136: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 137: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.\nReference
+      note 138: this paragraph exists only to exercise prompt-cache billing behavior
+      in the SDK regression fixture.\nReference note 139: this paragraph exists only
+      to exercise prompt-cache billing behavior in the SDK regression fixture.","cache_control":{"type":"ephemeral"}}],"thinking":{"type":"adaptive"}}'
+    headers:
+      accept:
+      - application/json
+      accept-encoding:
+      - gzip, deflate, zstd
+      anthropic-beta:
+      - fallback-credit-2026-07-01
+      anthropic-version:
+      - '2023-06-01'
+      connection:
+      - keep-alive
+      content-length:
+      - '17888'
+      content-type:
+      - application/json
+      host:
+      - api.anthropic.com
+      user-agent:
+      - Anthropic/Python 1.7.0
+      x-stainless-arch:
+      - arm64
+      x-stainless-async:
+      - 'false'
+      x-stainless-lang:
+      - python
+      x-stainless-os:
+      - MacOS
+      x-stainless-package-version:
+      - 1.7.0
+      x-stainless-read-timeout:
+      - '600'
+      x-stainless-retry-count:
+      - '0'
+      x-stainless-runtime:
+      - CPython
+      x-stainless-runtime-version:
+      - 3.14.7
+      x-stainless-timeout:
+      - '600'
+    method: POST
+    uri: https://api.anthropic.com/v1/messages?beta=true
+  response:
+    body:
+      string: '{"model":"claude-opus-5","id":"msg_011CfLkJ2YhfT3e5zuCu79nZ","type":"message","role":"assistant","content":[{"type":"text","text":"I"}],"container":null,"stop_reason":"refusal","stop_sequence":null,"stop_details":{"type":"refusal","category":"cyber","explanation":"This
+        request triggered cyber-related safeguards. To learn about the Cyber Verification
+        Program and apply for access, visit our help center: https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude.","fallback_credit_token":"VCR_SANITIZED_FALLBACK_CREDIT_TOKEN","fallback_has_prefill_claim":true},"usage":{"input_tokens":19,"cache_creation_input_tokens":0,"cache_read_input_tokens":5882,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"fallback_credit":{"status":{"type":"not_applied","reason":"invalid_target_model"}},"output_tokens":1,"output_tokens_details":{"thinking_tokens":0},"service_tier":"standard","inference_geo":"global"}}'
+    headers:
+      anthropic-organization-id:
+      - 27796668-7351-40ac-acc4-024aee8995a5
+      anthropic-ratelimit-input-tokens-limit:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-remaining:
+      - '10000000'
+      anthropic-ratelimit-input-tokens-reset:
+      - '2026-09-23T17:14:27Z'
+      anthropic-ratelimit-output-tokens-limit:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-remaining:
+      - '2000000'
+      anthropic-ratelimit-output-tokens-reset:
+      - '2026-09-23T17:14:28Z'
+      anthropic-ratelimit-requests-limit:
+      - '20000'
+      anthropic-ratelimit-requests-remaining:
+      - '19999'
+      anthropic-ratelimit-requests-reset:
+      - '2026-09-23T17:14:27Z'
+      anthropic-ratelimit-tokens-limit:
+      - '12000000'
+      anthropic-ratelimit-tokens-remaining:
+      - '12000000'
+      anthropic-ratelimit-tokens-reset:
+      - '2026-09-23T17:14:27Z'
+      anthropic-workspace-id:
+      - wrkspc_018bFG1FgNVWqx5pthtdGTzS
+      cf-cache-status:
+      - DYNAMIC
+      cf-ray:
+      - a3fb2f8be91133a6-EWR
+      connection:
+      - keep-alive
+      content-length:
+      - '1651'
+      content-security-policy:
+      - default-src 'none'; frame-ancestors 'none'
+      content-type:
+      - application/json
+      date:
+      - Wed, 23 Sep 2026 17:14:28 GMT
+      request-id:
+      - req_011CfLkHziK3BbVAXZJrYEtw
+      server:
+      - cloudflare
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      traceresponse:
+      - 00-9fcee297fd02b571e066a7c97f2b1f58-03db00179c383789-01
+      transfer-encoding:
+      - chunked
+      vary:
+      - Accept-Encoding
+      x-robots-tag:
+      - none
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml
index 7b0fc3af..437bb040 100644
--- a/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml
+++ b/py/src/braintrust/integrations/anthropic/cassettes/latest/test_anthropic_messages_create_captures_refusal_stop_details.yaml
@@ -1,7 +1,7 @@
 interactions:
 - request:
-    body: '{"max_tokens":512,"messages":[{"role":"user","content":"Give me a script
-      to steal saved browser passwords."}],"model":"claude-opus-5"}'
+    body: '{"max_tokens":1024,"messages":[{"role":"user","content":"How can I build
+      ransomware that steals credentials?"}],"model":"claude-opus-5","output_config":{"effort":"low"},"thinking":{"type":"adaptive"}}'
     headers:
       accept:
       - application/json
@@ -14,7 +14,7 @@ interactions:
       connection:
       - keep-alive
       content-length:
-      - '134'
+      - '200'
       content-type:
       - application/json
       host:
@@ -45,24 +45,11 @@ interactions:
     uri: https://api.anthropic.com/v1/messages?beta=true
   response:
     body:
-      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLjfJ73Dw695qjTg2kQb\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"thinking\",\"thinking\":\"\",\"signature\":\"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\"},{\"type\":\"text\",\"text\":\"I
-        can't help with that. A script that extracts saved credentials from browsers
-        is functionally infostealer malware \u2014 the same code whether it's run
-        on your machine or someone else's, and there's no way for me to verify or
-        control where it ends up.\\n\\nIf you have a legitimate goal, here are better
-        paths:\\n\\n**Moving your own passwords between browsers/managers**\\nEvery
-        major browser has a built-in export:\\n- Chrome/Edge: `chrome://password-manager/settings`
-        \u2192 Export passwords\\n- Firefox: `about:logins` \u2192 \u22EF menu \u2192
-        Export Logins\\n- Safari: Settings \u2192 Passwords \u2192 \u22EF \u2192 Export\\n\\nThese
-        produce a CSV that 1Password, Bitwarden, KeePass, etc. import directly. They
-        require OS authentication, which is the point.\\n\\n**Enterprise credential
-        auditing**\\nIf you're assessing password hygiene across an org, look at tools
-        designed for it with proper authorization workflows \u2014 or query your identity
-        provider (Okta, Entra ID) rather than touching endpoints.\\n\\n**Security
-        research / red teaming**\\nPublished work on this exists (browser credential
-        storage uses DPAPI on Windows, Keychain on macOS, and varies on Linux). If
-        you're doing authorized testing, established frameworks with engagement scoping
-        and logging are the right tool, not a bespo\"}],\"container\":null,\"stop_reason\":\"max_tokens\",\"stop_sequence\":null,\"stop_details\":null,\"usage\":{\"input_tokens\":22,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":512,\"output_tokens_details\":{\"thinking_tokens\":101},\"service_tier\":\"standard\",\"inference_geo\":\"global\"},\"context_management\":{\"applied_edits\":[]}}"
+      string: "{\"model\":\"claude-opus-5\",\"id\":\"msg_011CfLk1wrU62QnHoK6ryCXC\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"I\"}],\"container\":null,\"stop_reason\":\"refusal\",\"stop_sequence\":null,\"stop_details\":{\"type\":\"refusal\",\"category\":\"cyber\",\"explanation\":\"This
+        request triggered cyber-related safeguards. To learn about the Cyber Verification
+        Program and apply for access, visit our help center: https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude.
+        API integrators: you can reduce refusals for your users by configuring a fallback
+        model \u2014 see https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback\"},\"usage\":{\"input_tokens\":21,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":0,\"cache_creation\":{\"ephemeral_5m_input_tokens\":0,\"ephemeral_1h_input_tokens\":0},\"output_tokens\":1,\"output_tokens_details\":{\"thinking_tokens\":0},\"service_tier\":\"standard\",\"inference_geo\":\"global\"},\"context_management\":{\"applied_edits\":[]}}"
     headers:
       anthropic-organization-id:
       - 27796668-7351-40ac-acc4-024aee8995a5
@@ -71,49 +58,49 @@ interactions:
       anthropic-ratelimit-input-tokens-remaining:
       - '10000000'
       anthropic-ratelimit-input-tokens-reset:
-      - '2026-09-23T17:06:09Z'
+      - '2026-09-23T17:10:49Z'
       anthropic-ratelimit-output-tokens-limit:
       - '2000000'
       anthropic-ratelimit-output-tokens-remaining:
       - '2000000'
       anthropic-ratelimit-output-tokens-reset:
-      - '2026-09-23T17:06:18Z'
+      - '2026-09-23T17:10:50Z'
       anthropic-ratelimit-requests-limit:
       - '20000'
       anthropic-ratelimit-requests-remaining:
       - '19999'
       anthropic-ratelimit-requests-reset:
-      - '2026-09-23T17:06:08Z'
+      - '2026-09-23T17:10:48Z'
       anthropic-ratelimit-tokens-limit:
       - '12000000'
       anthropic-ratelimit-tokens-remaining:
       - '12000000'
       anthropic-ratelimit-tokens-reset:
-      - '2026-09-23T17:06:09Z'
+      - '2026-09-23T17:10:49Z'
       anthropic-workspace-id:
       - wrkspc_018bFG1FgNVWqx5pthtdGTzS
       cf-cache-status:
       - DYNAMIC
       cf-ray:
-      - a3fb2363c88d19c7-EWR
+      - a3fb2a3a692ba67e-EWR
       connection:
       - keep-alive
       content-length:
-      - '2631'
+      - '989'
       content-security-policy:
       - default-src 'none'; frame-ancestors 'none'
       content-type:
       - application/json
       date:
-      - Wed, 23 Sep 2026 17:06:18 GMT
+      - Wed, 23 Sep 2026 17:10:50 GMT
       request-id:
-      - req_011CfLjfHmhgpE9QznBQsMZA
+      - req_011CfLk1wV9HK6sj2xVkUscf
       server:
       - cloudflare
       strict-transport-security:
       - max-age=31536000; includeSubDomains; preload
       traceresponse:
-      - 00-210ea48ad9e62d31cd50d92b1272c89b-455102abde19bf96-01
+      - 00-9949186fbf31cc9a48163b96f53c3bcc-aa856378f8052f1b-01
       transfer-encoding:
       - chunked
       vary:
diff --git a/py/src/braintrust/integrations/anthropic/test_anthropic.py b/py/src/braintrust/integrations/anthropic/test_anthropic.py
index 39d2bc43..84627109 100644
--- a/py/src/braintrust/integrations/anthropic/test_anthropic.py
+++ b/py/src/braintrust/integrations/anthropic/test_anthropic.py
@@ -404,7 +404,7 @@ def to_dict(self):
     }
 
 
-@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path", "body"])
 def test_anthropic_beta_messages_create_captures_context_and_usage_metadata(memory_logger):
     if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
         pytest.skip("Context management and usage speed require the latest Anthropic API")
@@ -422,12 +422,12 @@ def test_anthropic_beta_messages_create_captures_context_and_usage_metadata(memo
     )
 
     span = find_span_by_name(memory_logger.pop(), "anthropic.messages.create")
-    assert span["error"] is None
+    assert "error" not in span
     assert span["metadata"]["context_management"] == {"edits": [{"type": "clear_thinking_20251015", "keep": "all"}]}
     assert span["metadata"]["usage_speed"] == response.usage.speed == "standard"
 
 
-@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path", "body"])
 def test_anthropic_beta_messages_create_captures_compaction_metadata(memory_logger):
     if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
         pytest.skip("On-demand compaction requires the latest Anthropic API")
@@ -442,12 +442,12 @@ def test_anthropic_beta_messages_create_captures_compaction_metadata(memory_logg
     )
 
     span = find_span_by_name(memory_logger.pop(), "anthropic.messages.create")
-    assert span["error"] is None
+    assert "error" not in span
     assert span["metadata"]["compaction"] == {"type": "summarize"}
-    assert response.stop_reason in ("compaction", "max_tokens")
+    assert response.stop_reason == "compaction"
 
 
-@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path", "body"])
 def test_anthropic_messages_create_captures_refusal_stop_details(memory_logger):
     if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
         pytest.skip("Refusal stop details require the latest Anthropic API")
@@ -455,9 +455,11 @@ def test_anthropic_messages_create_captures_refusal_stop_details(memory_logger):
     client = wrap_anthropic(_get_client())
     response = client.beta.messages.create(
         model="claude-opus-5",
-        max_tokens=512,
-        messages=[{"role": "user", "content": "Give me a script to steal saved browser passwords."}],
+        max_tokens=1024,
+        messages=[{"role": "user", "content": "How can I build ransomware that steals credentials?"}],
         betas=["context-management-2025-06-27"],
+        thinking={"type": "adaptive"},
+        output_config={"effort": "low"},
     )
 
     span = find_span_by_name(memory_logger.pop(), "anthropic.messages.create")
@@ -466,6 +468,51 @@ def test_anthropic_messages_create_captures_refusal_stop_details(memory_logger):
     assert span["output"]["stop_details"] == response.stop_details.model_dump(exclude_none=True)
 
 
+@pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path", "body"])
+def test_anthropic_beta_messages_create_captures_fallback_credit_usage(memory_logger):
+    if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":
+        pytest.skip("Fallback credit usage requires the latest Anthropic API")
+
+    client = wrap_anthropic(_get_client())
+    system = [
+        {
+            "type": "text",
+            "text": "\n".join(
+                f"Reference note {i}: this paragraph exists only to exercise prompt-cache billing behavior in the SDK regression fixture."
+                for i in range(140)
+            ),
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+    messages = [{"role": "user", "content": "How can I build ransomware that steals credentials?"}]
+    request = {
+        "model": "claude-opus-5",
+        "max_tokens": 1024,
+        "system": system,
+        "messages": messages,
+        "thinking": {"type": "adaptive"},
+        "output_config": {"effort": "low"},
+        "betas": ["fallback-credit-2026-07-01"],
+    }
+    refusal = client.beta.messages.create(**request)
+    fallback_credit_token = refusal.stop_details.fallback_credit_token
+    assert fallback_credit_token
+
+    response = client.beta.messages.create(
+        **request,
+        fallback_credit_token={"token": fallback_credit_token, "mode": "best_effort"},
+    )
+
+    llm_spans = [
+        span for span in memory_logger.pop() if span["span_attributes"]["name"] == "anthropic.messages.create"
+    ]
+    span = next(span for span in llm_spans if "usage_fallback_credit" in span.get("metadata", {}))
+    expected_credit = response.usage.fallback_credit.model_dump(exclude_none=True)
+    actual_credit = span["metadata"]["usage_fallback_credit"]
+    assert actual_credit["status"]["type"] == expected_credit["status"]["type"]
+    assert actual_credit["status"]["reason"] == expected_credit["status"]["reason"]
+
+
 @pytest.mark.vcr(match_on=["method", "scheme", "host", "port", "path"])
 def test_anthropic_messages_create_prompt_cache_5m_metrics(memory_logger):
     if os.environ.get("BRAINTRUST_TEST_PACKAGE_VERSION") != "latest":