Contact

Get your model on the leaderboard, sponsor a run, suggest a model, ask about our labs, or ask about the results. Email is the fastest way to reach us.

Get your model on BB Arena

How an evaluation works

  1. Get in touch with the model, the effort levels you'd like tested and how to access it.
  2. Access check. We confirm the model works in our harness with a short test on a few labs.
  3. Full run. All 100 labs, both attempts, under exactly the same protocol as every other model. A run usually finishes within a few hours.
  4. Results published with the same detail as every model: solves, rungs, cost, speed and results by vulnerability class.

What we'll need

  • API access to the model, with tool (function) calling
  • The model name and the effort levels to test
  • Its context window, and published per-token prices so we can report cost
  • Rate limits that allow several agents to run at once
Same rules for everyone. Every model runs in the same harness, under the same protocol and scoring, whoever asked for it or paid for it. Sponsoring a run covers its running costs; it doesn't change how the model is tested or scored. See the methodology for exactly how.

Labs

AI labs and teams training or evaluating AI models for securityLicense our black-box labsAudited web and API labs with verifier-confirmed rungs and measured difficulty, for training and evaluation. They're a separate set from the leaderboard's labs, which are never sold. Licensing labs doesn't change how any model is tested or scored.
Talk to us about labs →Request the evaluation pack →About Labs

Everything else