CEO-Bench - Mean Survival Days: leaderboard
CEO-Bench (2026) puts an agent in charge of a simulated software startup for 500 days, writing code against 34 tools and a 19-table business database with 1 million dollars of starting cash. The score is the mean number of days survived before bankruptcy over three simulations, out of 500, with each model at its maximum reasoning effort. Higher is better.
Metric: Mean days survived before bankruptcy (out of 500) over three 500-day CEO-Bench simulations of a startup starting with $1M cash (seed 42), the agent writing code against 34 tools and a 19-table business database at the maximum reasoning effort the model provides; a run that avoids bankruptcy counts 500; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash (High) | 500 | #101 (Gemini 3.5 Flash) |
| 2 | Qwen 3.7 Max (Thinking) | 500 | #145 (Qwen 3.7 Max) |
| 3 | Claude Sonnet 5 (Max) | 470.7 | #111 (Claude Sonnet 5) |
| 4 | Claude Fable 5 (Max) | 461.7 | #27 (Claude Fable 5) |
| 5 | Claude Sonnet 4.6 (Max) | 419.3 | #161 (Claude Sonnet 4.6) |
| 6 | Claude Opus 4.7 (Max) | 414.3 | #79 (Claude Opus 4.7) |
| 7 | GPT-5.5 (xHigh) | 401 | #49 (GPT-5.5) |
| 8 | Claude Opus 4.8 (Max) | 378 | #56 (Claude Opus 4.8) |
| 9 | GLM-5.2 (Max) | 376.7 | #138 (GLM-5.2) |
| 10 | GPT-5.6 Sol (Max) | 294 | #30 (GPT-5.6 Sol) |
| 11 | Claude Haiku 4.5 (Thinking) | 286.3 | #568 (Claude Haiku 4.5) |
| 12 | Gemini 3 Flash (High) | 137 | #164 (Gemini 3 Flash) |
| 13 | DeepSeek V4 Pro (Max) | 136.3 | #188 (DeepSeek V4 Pro) |
| 14 | Grok 4.20 (Reasoning) | 36 | #297 (Grok 4.20) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ceo-bench-mean-survival-days · How It Works · Data refreshed daily, snapshot 2026-11-13.