Agentic Task Completion on Terminal-Bench Hard 2 (30 tasks)
56.7Pass@1Codex
Evaluation results
| Method | Links | |
|---|---|---|
| CodexBase Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 56.7 | |
| TF-GRPOBase Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 55.6 | |
| AHEBase Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 53.3 | |
| NexAU0Base Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 51.7 | |
| ACEBase Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 48.9 | |
| terminus-2Base Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 40 | |
| opencodeBase Model=GPT-5.4 (high reasoning), Rollouts (k)=22026.04 | 33.3 |