llm-anthropic 0.25 adds Claude Opus 4.7 support with xhigh thinking effort

Simon Willison’s llm-anthropic plugin bumped to 0.25, wiring up claude-opus-4-7 as a new target model. The release introduces an xhigh tier for the thinking_effort parameter, extending the range of controllable reasoning budgets available through the CLI.

Two new boolean options land alongside the model: thinking_display and thinking_adaptive. The summarized thinking output surfaced by thinking_display is currently constrained to JSON output modes and JSON logs rather than the default text stream.

Housekeeping changes round out the release. Default max_tokens now tracks each model’s maximum allowed ceiling instead of a conservative static value, and the plugin drops the obsolete structured-outputs-2025-11-13 beta header that older models no longer require.