v1.18.0

Added

• Anthropic: cataloged claude-opus-5-5, claude-sonnet-5, and claude-fable-5-1 alongside the existing Opus 4.8 / Sonnet 4.6 / Haiku 4.5 models, with pricing and 1M-token context windows.
• OpenAI: cataloged gpt-6-astra, gpt-6-sol, and gpt-6-luna — Responses-API-only reasoning models with a 1,050,000-token context window — alongside the existing GPT-4o / GPT-5.5 family.
• Gemini: cataloged gemini-3.8-flash and gemini-3.5-flash-lite alongside the existing Pro 3.1 / Flash 3.5 / Flash-Lite 3.1 models.
• Setup wizard shows list price per model. Each entry in the commitbrief setup model picker reads <model> · $X/1M in · $Y/1M out · not measured, and the provider’s default model is marked default. Prices come from the provider catalog; quality is reported as “not measured” until benchmark results ship. See Providers.
• Eval corpus grown to 40 fixtures — 20 clean controls and 20 buggy — across Go, Python, JavaScript, and TypeScript. The live eval can repeat the corpus several times per run and report the per-run spread, and it can write a machine-readable results row (recall, precision, false-alarm rate on clean fixtures) to a JSON file. Pass/fail thresholds are fixed in code and cannot be relaxed per run.

Changed

• Default models (behavior change). A fresh config, and any config that leaves providers.<name>.model unset, now reviews with a different model on two providers. None of these defaults was chosen from a quality measurement, and none carries a quality claim. Configs that pin a model explicitly are unaffected; pin the old model to keep the old behavior.
  • gemini: gemini-3.8-flash (was gemini-3.5-flash). Chosen by list price: the cheapest model in Google’s current mid tier whose typical-diff estimate stays under the default $0.50 cost warning. A typical staged diff (about 8,500 input tokens) estimates at about $0.012, rising to about $0.024 when Google’s scheduled price increase takes effect on 2027-01-01.
  • anthropic: claude-opus-5-5 (was claude-opus-4-8), the current flagship. A typical staged diff estimates at about $0.064, but the model thinks by default and the cost estimate does not count thinking tokens, so the billed amount can be higher: in the worst case (output up to the 16,000-token ceiling plus one JSON repair retry) a single review can reach about $0.88 without the cost prompt appearing.
  • openai: unchanged, gpt-5.4-mini.
• Gemini: cached-input prices for all cataloged models now match Google’s published context-caching read price, replacing an earlier approximation.

Fixed

• commitbrief setup: pressing Enter on the model picker now accepts the provider’s default model instead of the first model in its list (Anthropic, OpenAI, Gemini, DeepSeek, Mistral, Cohere; Ollama’s discovered list is unchanged).
• The live eval now sends the model the same line-numbered diff a real review does, so eval measurements score the prompt the CLI actually ships.