release: 3.40.0 — surface the models 3.39.0 priced but never exposed - #127
Merged
Conversation
added 3 commits
August 19, 2026 20:54
3.39.0 added eleven chat models to src/pricing.ts and stopped there: no shortcut, no picker row, no router wiring. They were live, correctly priced, and reachable only by typing the full provider/model id. - 38 new shortcuts (107 -> 145): the GPT-5.6 pro reasoning tier, GPT-5.5 Pro, chat-latest, Gemini 3.6 Flash + the Flash Lite line, Qwen3.7 Plus/Flash, Tencent HY3, Xiaomi MiMo, and the 4o / 4.1-mini / o3-mini gaps - zai/glm-5.3 (new flagship, `glm` follows it) and xai/grok-imagine-video-1.5 - fix(router): the LLM classifier has been dead since 2026-07-27 — its default model was NVIDIA's EOL'd qwen3-next, which the gateway rides on nemotron-3-super-120b; that substitute leaks prose, so every classification failed the strict parse and fell through to keyword-only routing. Live-probed the free pool on the real classifier prompt: only nemotron-3-nano-omni returns one bare word. It is the new classifier. - nemotron-3-nano-omni serves itself again (re-probed on both the streaming and non-streaming paths) — it gets an alias, a free picker row, and second place in every free chain; mistral-nemotron slides to third. step-3.7-flash stays deliberately unaliased: it is still served by a pooled substitute. - picker holds its 24-row cap — GLM-5.3, Terra Pro and Qwen3.7 Flash displaced GLM-5.2, the hidden grok-fast row and GLM-5; every retired row keeps its shortcut, and `gemini` now tracks the flagship like `gpt`/`grok`/`glm` - brand numbers re-synced (95 visible, 8 video, 5 free); stale offline picker prices corrected 646 local tests pass.
Surfacing the catalog turned up three tables that had drifted behind it. - fix(vision): xai/grok-4.5 was missing from the vision allowlist. It is what `grok` resolves to and it IS vision-capable, so every image turn on the xAI flagship was rerouted to a vision sibling the user never chose. Eleven more were missing (GPT-5.6 pro tier, chat-latest, gpt-5.3, gpt-4o, Gemini 3.6 Flash + 3 Flash Preview, nemotron-3-nano-omni). New test pins the invariant: every bare flagship alias must be vision-listed. - fix(tokens): grok-4.5 (500K), grok-4.3 (1M) and glm-5.3 (1M) had no context window entry and match no inference pattern, so a cold catalog cache used the blind 128k default and compacted up to 8x too early. 12 entries added. - fix(context): the weak-model guidance branch matched bare `glm`, from the GLM-4.x era. `glm` now resolves to GLM-5.3 — 1M context, always-on reasoning — and was being told to make ONE tool call and stay under 300 words. GLM-5.x moves to balanced guidance. - fix(brain): the third extraction fallback, nvidia/nemotron-super-49b, left the catalog; use the current free default. 647 local tests pass.
zai/glm-5.3 is BASE-ONLY — sol.blockrun.ai lists 92 of Base's 93 models and 5.3 is the one it's missing (verified 2026-08-20; prices are otherwise identical across chains). Pointing the bare `glm` alias at it, as the flagship-tracking convention would, hands every Solana user an HTTP 400 on `/model glm`. Bare aliases have to resolve on both chains — they're what people type from muscle memory. `glm`/`glm5` stay on 5.2 (both chains, same price, one generation back); 5.3 keeps its explicit alias and its picker row, which reconcilePicker drops for Solana users on its own. 647 local tests pass.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
3.39.0 re-synced pricing to three weeks of gateway drift and added eleven chat models to
src/pricing.ts— then stopped. None of them got a shortcut, a picker row, or router wiring, so they were live, correctly priced, and reachable only by typing the fullprovider/modelid from memory.What
sol-pro/terra-pro/luna-pro), GPT-5.5 Pro,chatgpt(chat-latest), Gemini 3.6 Flash + the Flash Lite line, Qwen3.7 Plus/Flash, Tencent HY3, Xiaomi MiMo, plus the 4o / 4.1-mini / o3-mini gaps. Every alias target has a pricing entry and was probe-verified against the live gateway.zai/glm-5.3(Z.AI flagship, 1M ctx,glmfollows it) andxai/grok-imagine-video-1.5($0.08/s) in VideoGen.nvidia/qwen3-next-80b-a3b-instruct— NVIDIA's EOL'd id, which the gateway rides onnemotron-3-super-120b. That substitute opens with "Okay, let's see. The user wants to…" and never reaches a verdict inside the 16-token budget, so every classification failed the strict parse and fell through to keyword-only routing. Re-probed the free pool on the real classifier prompt — onlynemotron-3-nano-omnireturns one bare word (MEDIUM). It is the new classifier.gpt-oss-120bin August; re-probed 2026-08-19 on both the streaming and non-streaming paths it answers as itself. Getsomni, a free picker row, and second place in every free chain; degradedmistral-nemotronslides to third.nvidia/step-3.7-flashstays deliberately unaliased — $0 and in the catalog, but still served by a substitute.grok-fastrow and GLM-5; Gemini 2.5 Pro retires from under 3.1 Pro. Every retired row keeps its shortcut, andgemininow tracks the flagship likegpt/grok/glm.Verification
npm run build && npm test→ 646 pass, 0 fail (5 expectation updates, each documented in-line: the context example list, the free-profile allowlist, the picker bounds, and twopickFreeFallbackchain assertions).opus-4.6,gpt-5-nano,grok-3,grok-4-0709,grok-4-1-fast-reasoning) all still return 402 and stay.🤖 Generated with Claude Code