TERMS-Bench - Surplus Efficiency: leaderboard

Surplus efficiency on TERMS-Bench (2026), a bilateral price-negotiation benchmark in which LLM agents bargain against simulated counterparts with hidden reservation prices, urgency and stance across six behaviour families. The score is the utility the agent earns divided by the width of the zone of possible agreement, averaged over feasible episodes: 1 is the whole surplus, a failed deal counts 0 and a loss-making deal counts below 0. Higher is better.

Metric: Surplus efficiency on feasible episodes (SE+, maximum 1; loss-making agreements are not clipped and count below 0): the agent's utility divided by the zone-of-possible-agreement width, averaged over 1,800 seeded episodes per agent (six counterpart families, three regimes), deterministic decoding; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.6Max0.694#113
2GLM-5.10.686#198
3Claude Opus 4.7Max0.66#85
4Gemma 4 31B (IT)Thinking0.64#350
5Gemini 3.1 Pro (Preview)High0.639#99
6DeepSeek V4 ProMax0.618#196
7GPT-5.5xHigh0.606#54
8Qwen 3.6 Plus0.604#238
9Grok 4.20Reasoning0.601#301
10Kimi K2.60.597#192
11GPT-5.4xHigh0.531#135
12Seed 2.0 ProHigh0.522#222
13GPT-4o Mini0.189#1667

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=terms-bench-surplus-efficiency · How It Works · Data refreshed daily, snapshot 2026-11-21.