[HTExplicit]

Problem

OpenAI-format requests forwarded to Claude models other than Opus/Sonnet 5.5 mapped every reasoning effort the same way. This covers Responses and Chat Completions, including the native Anthropic-protocol paths. The single mapping was xhigh → max, plus thinking {type: enabled, budget_tokens} for every effort except low, with a default max_tokens of 8192.

Per the official Messages contract (checked 2026-10-02), that produces 400s:

Models	Rejected because
Opus 4.7/4.8/5, Sonnet 5, Fable/Mythos 5.x	budget_tokens
Opus 4.6, Sonnet 4.6, Opus 4.5, Mythos Preview at high and above	budget ≥ max_tokens
Opus 4.5	max is not a valid level
Haiku/Sonnet 4.5	output_config.effort not supported
none/minimal on every non-5.5 model	invalid effort sent

A read-only check of production found no such traffic in the last 30 days, so the bug is latent.

Change

One conversion (applyReasoningEffort) is driven by the effort catalog for the resolved upstream model. All call sites use it: ForwardAsResponses, ForwardAsChatCompletions, and the two native Anthropic-protocol paths.

• Adaptive families:
  • thinking: adaptive at the requested level, or the nearest level the family lists (deeper first). xhigh stays native where listed and becomes max on Opus 4.6 / Sonnet 4.6.
  • none/minimal → lowest level with adaptive thinking. Anthropic recommends this over disabling thinking on Opus 5 and 4.8.
• Budget families and Claude models without effort levels:
  • Budget thinking from medium up, with 1024 ≤ budget < max_tokens guaranteed. A client max_output_tokens is kept and the budget shrinks to fit; the default max_tokens grows by the budget, capped at 32000.
  • No budget thinking when a tool call is forced.
  • No output_config.effort for models without levels.
• Sampling: a request that thinks drops temperature ≠ 1 and top_p < 0.95.
• Model IDs: dotted OpenRouter IDs (claude-opus-4.7) normalise to their catalog family. The 5.5 checks stay exact.
• Unchanged: the 5.5 branch and non-Claude upstreams (GLM/Kimi/DeepSeek behind Anthropic-compatible endpoints).
• Usage log and billing read the effort that is actually forwarded.

Known and deferred: the Antigravity transformer ignores output_config.effort for native and compat traffic alike. There are 0 Antigravity accounts in production.

Verification

• TestResponsesToAnthropicReasoningEffortByModelFamily (table across families, including sampling and forced-tool cases, plus the Chat Completions chain);
• the 5.5 regression tests;
• the effort catalog tests;
• the compat pricing test;
• gofmt, go build ./..., and go vet on the changed packages.

Not run: real model calls.

🤖 Generated with Claude Code