Newshot Leaderboard
*Every model runs at max effort. Gemini 4 Argon runs at high.
†Most chambers are still being run, due to compute constraints and usage limits.
Currently running!
Cost per Chamber
Plans per Chamber
Time per Chamber
Results
Changelog
- SoonResearch writeup
- Nov 9, 2026Added results at max effort for GPT-6.1 Sol, GPT-6 Astra and GPT-6 Luna
- Nov 3, 2026🎉 Portal2Bench announced, with results at max effort for Opus 5.5, Sonnet 5.5, Fable 5.1 and GPT-6.1 Terra
Yes. GPT-6.1 Sol (Codex) solves 20 of 26 Portal 2 test chambers. GPT-6 Astra (Codex) solves 20, GPT-6.1 Terra (Codex) 19, Opus 5.5 (Claude Code) 11, Sonnet 5.5 (Claude Code) 6, GPT-6 Luna (Codex) 0 and Fable 5.1 (Claude Code) 3 of 15 so far.
Portal2Bench pauses Portal 2 while the model thinks. The agent sends keys and view turns, and gets back screenshots and its position. It has a notes folder, but no shell, web or console. Each chamber is a fresh run, solved if the agent reaches the next map within 300 plans and $30. 26 chambers from chapters 2 to 5 count. Each model runs in its own harness at its highest effort.
Portal2Bench prices each run at API list prices. A solved chamber's median cost runs from $0.74 (GPT-6.1 Terra) to $13.12 (Fable 5.1).
No. The agent has no shell or web, and its file tools reach only its own notes folder. A QA check flags anything else.
The agent respawns at its last checkpoint, as in normal play, and is told only that it died.
No. Portal2Bench is not affiliated with or endorsed by Valve.