The update also introduces thinking_display and thinking_adaptive options, increases default max_tokens to each model’s limit, and removes an obsolete structured-outputs header.
AI Quick Take
- claude-opus-4.
- Default max_tokens raised to model limits and the older structured-outputs-2025-11-13 beta header has been removed for legacy models.
Simon Willison released llm-anthropic 0.25, which adds a new model, claude-opus-4.7, and introduces support for a higher internal effort setting via thinking_effort: xhigh.
The package also adds two boolean flags, thinking_display and thinking_adaptive. Summarized output produced by thinking_display is currently exposed only in JSON output or JSON logs. Separately, the release increases the default max_tokens setting to each model’s maximum allowed value and removes the older structured-outputs-2025-11-13 beta header previously used for legacy models.
Operationally, these are configuration and compatibility changes: the new thinking flags formalize how callers can request and receive internal 'thinking' summaries, but those summaries are limited to JSON for now. The raised max_tokens simplifies usage where full context is expected, and eliminating the obsolete header reduces beta-era legacy behavior. Teams should watch for further SDK documentation and any expansion of thinking_display beyond JSON outputs.