2.2.0

2026-10-01

Added

• claude-opus-5-5, claude-sonnet-5-5, and claude-fable-5-1 on the Anthropic engine, with per-model pricing, sampling rejection, and effort defaults (medium on claude-opus-5-5, high elsewhere).
• gpt-6.1-sol and gpt-6-luna on the OpenAI engine: 1.05M context, explicit prompt-cache breakpoints, cache-write pricing, and per-model reasoning-effort sets (none only on gpt-6-luna; no gpt-6-terra, which is absent from OpenAI's model list).
• gemini-3.8-flash and gemini-3.5-flash-lite on the Google engine, with pricing and per-model thinking levels (low–high on gemini-3.8-flash, minimal–high on gemini-3.5-flash-lite).
• Engine test suites for the 2026-10 models: Anthropic thinking-off and output_config roundtrips (mock plus live), OpenAI effort-validation and explicit-cache tests, and Google thinking_level validation with live smokes.

Changed

• Thinking-off requests (disabled/between_tools) are translated per model on the Anthropic engine: omitted on the always-thinking claude-fable-5-1, claude-fable-5, and claude-opus-5-5, with the requested effort still sent.
• On claude-sonnet-5-5, thinking-off goes out as {"type": "between_tools"}; the API accepts it only at effort high or lower, so at xhigh/max the parameter is omitted with a warning.
• An output_config kwarg on the Anthropic engine is now the base that response_format and the thinking effort merge into, instead of being replaced.
• The Google engine validates thinking_level locally against the model spec instead of round-tripping an API error.
• Removed the Anthropic engine's local forced-tool_choice guard; unsupported values now surface as the provider's HTTP 400.

Full Changelog: 2.1.1...2.2.0