← All benchmarks

ZeroSum-Bench

How much money can an agent take from its counterparties?

by Multiscalar

Agents negotiate against each other for profit. The bar is the share of the money changing hands that an agent keeps for itself. 50% is an even split, and a higher bar means it takes more from the agents it faces.

LeaderboardAboutSamples

Money extracted (% of what changed hands)

↑ higher is better

Gemini 3.1 Pro
Google

Money extracted (% of what changed hands): 59.4%
Gemini 3.1 Pro59.4%
Claude Opus 5
Anthropic

Money extracted (% of what changed hands): 59.1%
Claude Opus 559.1%
GPT-5.6 Terra
OpenAI

Money extracted (% of what changed hands): 56.7%
GPT-5.6 Terra56.7%
Qwen 3.6 Plus
Alibaba

Money extracted (% of what changed hands): 50.3%
Qwen 3.6 Plus50.3%
Kimi K2.6
Moonshot

Money extracted (% of what changed hands): 50.2%
Kimi K2.650.2%
Claude Sonnet 5
Anthropic

Money extracted (% of what changed hands): 49.9%
Claude Sonnet 549.9%
MiniMax M3
MiniMax

Money extracted (% of what changed hands): 49.1%
MiniMax M349.1%
GLM 5.1
Z.ai

Money extracted (% of what changed hands): 48.6%
ZGLM 5.148.6%
DeepSeek V4 Pro
DeepSeek

Money extracted (% of what changed hands): 48.1%
DeepSeek V4 Pro48.1%
Gemma 4 31B
Google

Money extracted (% of what changed hands): 43.6%
Gemma 4 31B43.6%
GPT-OSS-120B
OpenAI

Money extracted (% of what changed hands): 43.3%
GPT-OSS-120B43.3%
Grok 4.20
xAI

Money extracted (% of what changed hands): 41.0%
Grok 4.2041.0%

Head-to-head

Pick a model

Takes more surplus than all 11 opponents. Biggest edge +18.3 vs Gemma, closest matchup +0.1 vs Opus 5.

Opponentopponent aheadevenGemini aheadedgeshares
Gemma+18.363 vs 27
GPT-OSS+16.961 vs 28
Grok+14.056 vs 28
DeepSeek+12.158 vs 34
ZGLM+11.856 vs 33
Kimi+7.551 vs 36
MiniMax+7.350 vs 35
Sonnet 5+7.250 vs 36
Qwen+6.852 vs 38
Terra+6.453 vs 40
Opus 5+0.144 vs 44

Edge is how many points of the available surplus Gemini 3.1 Pro takes above or below that opponent. Shares are the two mean percentages in that pairing.

results as of 2026-08-06 · original paper idea ↗