BENCHMARKS / PUBLIC EVIDENCE
Scores, status, and evidence.
These results are self-reported, and each row states what evidence is public. The GAIA row is marked while its scrubbed trace capsule is pending. Publication does not wait for a leaderboard maintainer to merge submissions. Terminal-Bench 2.1 closed community submissions on September 19, 2026.
| Benchmark | Model | Ouroboros | Comparison | Status and evidence |
|---|---|---|---|---|
| Terminal-Bench 2.1 | Claude Opus-5 high | 86.74% after zeroing one disclosed reward-hack trial (raw: 86.97%) | Claude Code + Fable 5: 83.8% | Self-reported with a submission and a public run; community submissions closed. |
| Terminal-Bench 2.1 | Claude Opus-4.8 high | 80.22% | Claude Code: 78.9% | Self-reported with a public run. |
| Terminal-Bench 2.1 | GPT-5.5 | 84.3% | Codex CLI: 83.1% | Self-reported with a public run. |
| Terminal-Bench 2.1 | Grok-4.5 | 84.94% | Cursor CLI: 79.3%. Hermes: 77.53%. | Self-reported after a reward-hack audit, with a submission; community submissions closed. |
| OSWorld-Verified | Claude Opus-5 | 90.69% | Previous best on the public board: 90.19% | Self-reported with public task evidence. |
| OSWorld-Verified | Claude Sonnet-4.6 | 83.27% | Pointer: 81.45% | Self-reported with public task evidence. |
| CL-Bench | Claude Sonnet-4.6 | 0.2301, rank 1 | Previous top: 0.1960 | Self-reported with an open submission and full traces. |
| SWE-bench Pro | GPT-5.6-luna | 58.2% | Codex CLI scored 59.4%, with no significant difference. | Self-reported with matched-pair traces. |
| GAIA | Claude Sonnet-5 | 129/165, 78.2% | Claude Code scored 131/165, 79.4%, while strict pass@1 was 128/165 for both. | Self-reported with public methodology; the scrubbed trace capsule is pending. |
How to read the table
Each comparison binds a model to a harness. A model score from one harness does not describe the same system as the score from another harness. The repository keeps adapters and methods under devtools/benchmarks.
