Codex Warriors
Type any task, and we'll auto-generate a benchmark, run it blind across frontier AI models, then have Gemini write harder follow-up tests from the failures and rerun them so you don't just see who scored best, you see ex
Core trick
Gemini generates harder follow-up test cases from each model's failed answers, then reruns the benchmark to map where models break.
Description
Type any task, and we'll auto-generate a benchmark, run it blind across frontier AI models, then have Gemini write harder follow-up tests from the failures and rerun them so you don't just see who scored best, you see exactly where each model breaks.
README excerpt
# benchmarkyourselflah Prompt-to-benchmark prototype built with Next.js and Convex. Users type one benchmark prompt, generate a 12-case draft, trim categories and question count, run it across public models, then get judged results, a shareable URL, and an evolving benchmark tree that proposes harder follow-up cases from observed model failures. ## Local setup 1. Copy `.env.example` to `.env.local`. 2. Fill `OPENROUTER_API_KEY`. All benchmark generation, evolution-tree generation, judging, model runs, and supplemental answer images are routed through OpenRouter. 3. Start Convex: ```bash pnpm convex:dev ``` 4. Start the app: ```bash pnpm dev ``` 5. Open `http://localhost:3000`. ## Commands ```bash pnpm dev pnpm lint pnpm typecheck pnpm test pnpm convex:dev ``` ## Notes - Models stay hidden until `OPENROUTER_API_KEY` is configured. - Curated model lineup now includes current OpenRouter leaderboard picks such as `anthropic/claude-sonnet-4.6`, `google/gemini-3-flash-preview`, `deepseek/deepseek-v3.2`, `minimax/minimax-m2.5`, and `moonshotai/kimi-k2.5`. - Each finished run now gets a Gemini-generated evolving benchmark tree that extracts weak spots and writes harder follow-up cases for the results page. - Each completed model answer now gets a Nano Banana visual summary using `OPENROUTER_IMAGE_MODEL` (default `google/gemini-3.1-flash-image-preview`) so users can scan results quickly. - Anonymous rate limits are enforced at the route layer: 5 draft generations/hour and 3