[stackbilt-admin]

Refs #112 (phase 2). Unreleased; no version bump in this PR.

Every value was checked against vendor primary docs on 2026-10-15:

• Cloudflare: each Workers AI model page, the pricing page, and the 2026-05-08 and 2025-09-18 deprecation posts.
• OpenAI: developers.openai.com models, pricing and deprecations pages.
• Anthropic: platform.claude.com models overview, pricing, deprecations and the 5.x migration guides.

No live or billable vendor requests were made.

Review history.

• Round 1: making the new Claude 5.x and GPT-5.6 models the defaults would break live calls. They are now explicit-only, the adapter request shapes are fixed, and a JSON-mode bug that already affects 1.22 is fixed.
• Round 3 (verdict "ready to merge"): follow-ups applied: the timeout note now mentions retries, the README states which Anthropic model each request shape gets, the gpt-4o-mini cost test runs through the factory and checks the ledger, and mutation counts cover every test file.
• Round 2 (verdict "merge after fixes"): no code blockers. This revision corrects false or missing release notes, restores the { in streamed JSON on prefilled models, and adds factory-level and streaming tests. It also opens Anthropic adapter sends tool_result blocks without tool_use_id, so multi-turn tool use fails #124: tool_result blocks are sent without tool_use_id, an older bug that affects every Claude model.

Cloudflare (primary provider)

• Retired in the catalog: the 12 models Cloudflare deprecated.
  • On 2026-05-30: Llama 3.1 8B and 70B, Llama 3 8B, Llama 2 7B int8, Phi-2 and Mistral 7B v0.1.
  • On 2025-10-01: OpenChat 3.5, TinyLlama and the four Qwen 1.5 models.
  • They were compatibility, so routing could still pick them. They now stay accepted by name but are never auto-routed.
• Context windows corrected:
  • Llama 3.3 70B FP8 Fast: 24,000 (was 131,072)
  • Llama 3.2 1B: 60,000
  • Llama 3.2 3B: 80,000
  • Qwen3 30B: 32,768
  • Kimi K2.6 and K2.7 Code: 262,144
  • DeepSeek V4 Flash: 1,048,576
• Kimi K2.7 Code gains vision.
• Qwen 2.5 Coder no longer claims function calling, because Cloudflare's model page doesn't list it.
• AI Gateway unified-catalog models:
  • deepseek/deepseek-v4-pro and anthropic/claude-opus-4.8 were priced at $0. They now cost $1.74/$3.48 and $5/$25 per million tokens.
  • Opus 4.8 no longer receives temperature or top_p, which Claude 4.7 and later reject.
• Added:
  • @cf/meta/llama-3.1-8b-instruct-fp8
  • @cf/ibm-granite/granite-4.0-h-micro
  • @cf/aisingapore/gemma-sea-lion-v4-27b-it (explicit-only)
• Ranking test: a Workers AI model must rank first for every use case when all six providers are configured and healthy.

New ModelCatalogEntry.explicitOnly

This is an optional, additive field. An explicit-only model is served when a request names it, but recommendations and automatic routing skip it.

It is set on the new OpenAI models, the Claude 5.x models and SEA-LION. OpenAI and Anthropic defaults stay where they were in 1.22:

• gpt-4o-mini for every OpenAI use case;
• for Anthropic: Haiku 4.5 for general and cost-effective requests, Sonnet 4.6 for tools, vision and long context, and Opus 4.6 for high-performance and research.

Routing that does change, because the data was corrected:

• Cloudflare long-context requests now pick GLM-5.3 Flash instead of DeepSeek V4 Flash, because DeepSeek's context window was corrected from 1.31M to 1.05M.
• Llama 3.1 8B FP8 and Granite 4.0 Micro join the recommendation lists, though neither ranks first.
• OpenAI's fallback pool is now just gpt-4o-mini.

OpenAI

• Shutting down 2026-10-23: gpt-4, gpt-4-turbo and gpt-3.5-turbo are retired and no longer auto-routed, but are still accepted by name.
• gpt-3.5-turbo-16k is no longer a valid ID and is also retired.
• gpt-4o-mini stays the default. getRoutingInfo() reports its 2027-01-14 shutdown; response metadata doesn't, because the model is still active.
• Added (explicit-only): gpt-6-astra, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna and gpt-5.5.
  • Requests to them send max_completion_tokens instead of max_tokens, and no temperature.
  • Dropping temperature is an inference; OpenAI's model pages don't say whether these models accept it.
• Cost fixes:
  • A response naming a dated snapshot was priced at $0. It is now priced by its alias.
    • This includes the default gpt-4o-mini, whose responses name gpt-4o-mini-2024-07-18. Ledgers and budgets will start showing real spend for it after upgrading.
    • Snapshot IDs without a date suffix are still unpriced.
  • Prompts over 272K tokens on the new models bill at 2x input and 1.5x output.

Anthropic

• Added (explicit-only): claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5.
  • Requests omit temperature.
  • max_tokens defaults to 16,000, because these models think by default and thinking counts against max_tokens.
  • Forced tool choice on Fable 5.1, Opus 5.5 and Sonnet 5.5 throws a ConfigurationError before any API call, since those models return 400 for it. Haiku 5.5 still accepts it.
• JSON mode fix: the adapter no longer sends an assistant prefill to Opus 4.6, Sonnet 4.6 or the 5.x models.
  • Anthropic's guide says Sonnet 4.6 and later reject prefill with a 400. That means 1.22 JSON mode on the default claude-sonnet-4-6 already fails today.
  • These models rely on the JSON system instruction instead.
  • The { is restored only when it was prefilled, on both the non-streaming and streaming paths. Streaming previously never restored it.
• Pricing: cache reads use per-model rates, and Haiku 5.5 prompts over 100K tokens bill at the 5x rate.
• Legacy models: Opus 4.6, Sonnet 4.6 and Haiku 4.5 stay active, with a "legacy" note in their descriptions.

Known limitations (still open on #112)

• Thinking blocks in tool loops (Anthropic adapter sends tool_result blocks without tool_use_id, so multi-turn tool use fails #124): Claude 5.x thinking blocks are not passed back with tool results, and tool_result blocks lack tool_use_id. This is why the 5.x models are explicit-only.
• Timeouts: the 30-second default timeout can cut off long non-streaming Fable 5.1 and Opus 5.5 responses at the new 16,000-token default. Timeouts are retried (maxRetries, default 3), so a request can take about two minutes before fallback, and each attempt may be billed. This is documented, not changed.
• Qwen 2.5 Coder: requests to it that include tools now throw a ConfigurationError before any call.
• Unsupported OpenAI models: gpt-6-luna and gpt-6.1-sol restrict tool calling on Chat Completions.
• Not modelled yet:
  • OpenAI cached-input pricing;
  • Cerebras, Groq and NVIDIA prices and IDs;
  • a custom-model escape hatch;
  • unknown-cost semantics.

Validation

• tsc --noEmit, full vitest (789 tests; model-refresh.test.ts has 125), build, test:package and test:worker-import all pass.

• Factory-level tests now cover:

  • named explicit-only models being served by their own provider;
  • unnamed requests never landing on an explicit-only model;
  • forced tool choice failing on both the generate and stream paths without any provider call;
  • toolChoice: 'none'.

• factory.test.ts and catalog-accuracy.test.ts are unchanged from main.

• Undoing each fix makes its tests fail (counts across all test files):

  Fix	Tests that fail without it
  explicitOnly exclusion	19 (14 in model-refresh, 2 in model-catalog, 3 in factory)
  Streamed { restore	1
  Streamed restore check	1
  Prefill check	6
  Brace-restore check	6
  Forced tool choice	5
  5.x max_tokens	2
  Snapshot pricing	2 (including the factory ledger test)
  OpenAI long-prompt rate	1
  Opus 4.8 sampling	1
  OpenAI reasoning parameters	6
  Anthropic temperature	4
  Cache-read rates	3
  Haiku 5.5 long-prompt rate	1

🤖 Generated with Claude Code