SDK Core 0.3.7

Fixed

• Drop temperature and top_p (with a one-time warning) for Claude models that reject them: Opus 4.7 and later, Sonnet 5, Opus 5 and Fable, including Bedrock and Vertex ids. A Python Agent with default settings failed with a 400 on all of them. New or unrecognised Claude models are treated as rejecting.
• Raise the default Anthropic max_tokens from 1024 to 16384 for those models (4096 for older ones), since thinking counts toward it; Opus 5 answers were cut off mid-sentence. The Anthropic and Bedrock request timeout default rises from 30 s to 10 minutes to match the OpenAI provider.
• OpenAI reasoning models are now every gpt-N with N ≥ 5 plus the o-series, from one shared rule.
• Bedrock cross-region inference profile ids (us.anthropic.…, eu., apac., global.) route to their model family instead of failing as unsupported.
• Claude Haiku accepts sampling parameters only up to 4.5 (Opus and Sonnet up to 4.6).

Added

• ReasoningEffort::None and ReasoningEffort::Low. gpt-6 accepts none/low/medium/high and rejects minimal; gpt-5 keeps minimal. Gemini maps none to minimal.