llm-anthropic 0.25

The meaningful detail here isn't the new model support but the architectural choice to default to per-model maximum token limits, which quietly shifts how developers interact with extended thinking without requiring them to manually tune context windows. That default behavior change will affect cost and latency profiles for anyone already running the plugin in production.

This release lands in the middle of an unusually active week for Anthropic tooling. The Verge's coverage of OpenAI's Codex upgrade framed it as a direct challenge to Claude Code, and Willison's plugin update is a small but concrete data point on the other side: the developer tooling around Claude is also moving, not standing still. The extended thinking display toggles are particularly relevant given that Claude Opus 4.7 is the model being added here, and Anthropic's simultaneous push into cybersecurity (Claude Mythos Preview, per The Verge) and hardware assistance (Schematik, per WIRED) suggests the company is deliberately expanding the surface area where Claude gets integrated. Willison's work tends to be a leading indicator of where serious developers are actually spending time.

Watch whether Willison or other prominent plugin maintainers add native support for Claude Mythos Preview within the next four to six weeks. If that happens quickly, it signals the cybersecurity model is getting real developer traction beyond Anthropic's own tooling.