Use-case preset
Code review assistant cost calculator
PR diff and style guide in; inline review comments out.
A code review assistant receives a PR diff plus a project style guide and emits inline comments flagging bugs, style violations, and security issues. Diffs with surrounding context commonly run 20–28k tokens; comments are concise — typically 10–30 tokens each — yielding the 90/10 input/output split and 32k context ceiling.
The `under_5s_p95` target keeps review feedback snappy enough to land before the developer context-switches. cachedPromptPercent of 40 covers the stable system prompt, linting rules, and style guide prefix shared across every PR. Coding-specialist models (Codestral, Qwen-2.5-Coder, DeepSeek-Coder) consistently outperform general instruction models on diff comprehension at equivalent parameter counts. Quantized 32B coding models typically hit 90%+ of 70B accuracy at 40% lower cost — worth A/B testing in high-volume pipelines.