nagi-studio/nagi-bench

NAGI BENCH is a public gallery of AI model evaluation cases created by NAGI STUDIO. The idea is simple: take the same creative prompt and run it through many different AI models and tools, collect the outputs as working HTML or SVG files, and display them side by side so anyone can compare the results. A key concept in this project is the distinction between a model and a harness. A model is the underlying AI engine, such as GPT-5.5 or Claude Fable 5. A harness is the product or tool that wraps the model and controls how it receives instructions, manages conversation history, calls external tools, and decides when to stop. Two entries that use the same model but different harnesses are treated as different agents, because the harness often has as much influence on the output as the model itself. The collection currently covers models from Anthropic, OpenAI, Google, xAI, and several Chinese AI companies, each run through tools like Claude Code, Cursor, Codex CLI, or web chat interfaces. The two evaluation prompts in use right now ask models to generate a playable HTML world with fictional creatures, and to draw a pelican cycling by the sea as an SVG image. Both prompts are kept fixed so that every submission is answerable from context alone, without external tools or files. The site at bench.nagi.fun includes a blind voting mode where two random outputs are shown anonymously and visitors pick the better one before the model names are revealed. Once enough votes accumulate, a leaderboard is generated using the same ranking method used by LMSYS Chatbot Arena. Votes are stored locally in the browser by default, or in a Cloudflare backend if the project operator sets one up. Contributing a new result means adding the output file and a short JSON registration entry for the model and harness combination. No code changes are needed. A CI script checks that all entries are consistent before any submission is accepted.