Benchmark report
Methodology and per-scenario results for Codna vs. leading coding agents.
Frequently asked
Codna was tested head-to-head against Cursor. It used 5× fewer tokens and ran 1.7× faster, with every fix test-verified.
Token consumption, wall-clock speed, and verified fix rate across 8 real bug-fix scenarios run against OpenAI Codex CLI and Google Gemini CLI. Every fix counted only when your own tests passed.
A deterministic engine maps the repository in roughly 60ms without any LLM calls. It then hands the AI agent an evidence bundle of around 600 tokens — measured 162x smaller than reading the full repo — so the agent fixes the right code immediately.
Yes. The engine has mapped 130 repositories in 9.2 seconds, consuming zero tokens for the mapping step.
Codna locates the affected code using its dependency and blast-radius graph, generates a fix from the evidence bundle, then runs your tests to verify the result. On GitHub, it opens a pull request with the verified fix.
You can self-host the engine, bring your own API key, and configure fail-closed egress. Codna never trains on your code.