Skip to content

v0.78.0

Latest

Choose a tag to compare

@github-actions github-actions released this 18 Nov 23:20

Changed

  • Breaking: AnthropicModelCapabilities.SupportsDisabledThinking (bool) is replaced by ThinkingOffStyle (AnthropicThinkingOffStyle: LowEffort, Disabled, BetweenTools). How Claude turns thinking off differs per model, so a bool could not say it: Sonnet 5 and Opus 4.6-4.8 take thinking: {type: "disabled"}, Sonnet 5.5 takes {type: "between_tools"}, and Fable, Opus 5 and Opus 5.5 take the lowest effort instead. Migration: an override that set SupportsDisabledThinking = true becomes ThinkingOffStyle = AnthropicThinkingOffStyle.Disabled; false is the default (LowEffort) and needs no line. An override is matched by longest prefix like the built-in rows, so one keyed claude-sonnet-5 also covers claude-sonnet-5-5 and hides its built-in row: give claude-sonnet-5-5 its own entry (BetweenTools) or key the override by the exact id.

Fixed

  • MessageThinkingEffort.Minimal and None on the Gemini 4 models (gemini-4-argon, gemini-4-carbon, gemini-4-hydrogen) no longer ask for thinkingLevel: minimal, and Hydrogen no longer receives temperature, topP or topK. The vendor lists low, medium and high for them (xhigh too on Hydrogen), says minimal is «Not supported on Gemini 4 models», and that Hydrogen «rejects minimal with an error» and wants the sampling parameters removed; the ids had no row and fell to the default, which accepts minimal and forwards sampling values. They now leave as low, and Hydrogen drops the sampling values as Gemini 3.8 Flash does (measured: the server accepts and ignores in-range values and answers 400 to an out-of-range one, so dropping them is the vendor's advice rather than a cure for a failure). A MessageThinkingEffort.XHigh request still goes out as high: Hydrogen accepts xhigh (the other Gemini models answer 400), but Google.GenAI 2.0.0 has no ThinkingLevel value for it. Measured against the API on 2026-11-19: minimal is 400 («Thinking level MINIMAL is not supported for this model») on all three, low is 200.
  • MessageThinkingEffort.None on gpt-6-astra, gpt-6.1-sol, gpt-6.2-sol and gpt-6.2-astra no longer asks for reasoning.effort: none. These models take low to max only (the vendor's GPT-6.2 page: «The API rejects requests that set reasoning effort to none»), but the gpt-6 ids had no row and fell to the default, which includes none. They now send low, the lowest value they accept; gpt-6-sol, gpt-6-luna and gpt-6.1-terra list none and are unchanged. Measured against /v1/responses on 2026-11-19: none is 400 on the four («Supported values are: 'low', 'medium', 'high', 'xhigh', and 'max'») and 200 on the three.
  • A required or named tool choice on Claude Sonnet 5.5, Fable 5.2 or Mythos 5.2 no longer fails with HTTP 400. These models reject tool_choice any and tool like Fable 5.1 and Opus 5.5 do, but had no row of their own (Sonnet 5.5 matched the Sonnet 5 row, the 5.2 ids matched none), so the request carried the forced value. It now leaves as auto plus the explicit tool-call instruction, as for the other models that reject it. Claude Haiku 5.5 accepts a forced choice and is unchanged.
  • MessageThinkingEffort.None on claude-sonnet-5-5 no longer fails every request with HTTP 400. The model was matched to the Sonnet 5 row and sent thinking: {type: "disabled"}, which Sonnet 5.5 refuses; it now has a row of its own and sends thinking: {type: "between_tools"} together with the lowest effort (output_config.effort: low, as for the other models that cannot switch thinking off; between_tools takes up to high, answers 400 to xhigh / max, and takes no display or budget). between_tools turns off the thinking before a reply; thinking between tool calls remains. Sonnet 5 keeps disabled; a request with a thinking level other than None is unchanged. Measured against the API on 2026-11-19: disabled 400; between_tools alone 200, with low / medium / high 200, with xhigh / max 400, with display or budget_tokens 400 («Extra inputs are not permitted»).