TERMS-Bench: leaderboard
Mean native utility over 1,800 bilateral price-negotiation episodes: 600 each for overlap, urgency-shift and no-deal regimes. Reasoning models use the paper’s declared maximum reasoning setting. Raw utility is scale-dependent; it is not surplus efficiency or a percentage.
Metric: Mean Utility. Source: terms-bench.github.io. Saturation forecast: Around September 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (xHigh) | 11.09 | #49 (Claude Opus 4.7) |
| 2 | GPT-5.5 (xHigh) | 10.15 | #31 (GPT-5.5) |
| 3 | GPT-4o Mini | 3.37 | #599 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=terms-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.