Atlas

Benchmarks

← All benchmarks

YC-Bench

Games · 2026-04-01

Collinear AI's long-horizon business simulation: an agent runs a simulated AI startup for one year through a CLI, starting with $200,000, choosing task contracts, assigning eight employees, building client trust and avoiding adversarial clients while payroll grows, with only a 20-round context and a persistent scratchpad. The leaderboard score is final net worth in USD on the default preset, averaged over three seeded worlds with a bankrupt run counted as zero; it is unbounded and heavy-tailed.

Top models (higher is better)

ModelScore
Claude Opus 5.52873725
Grok 4.52629903
Qwen3.8-Max2095838
Kimi K32053296
Claude Fable 51977573
Claude Opus 51911592
MiMo-V2.6-Pro1881623
Claude Sonnet 5.51741315
Opus 4.71714511
Opus 4.81711389
Grok 4.71666389
DeepSeek V4.1 Flash1657236
GLM-5.31597182
Grok 4.61568241
GLM-5.11510772
DeepSeek V4 Pro 08131459743
GPT-6 Sol1441793
Qwen3.8-Omni-Flash1380061
GPT-6 Astra1336466
Opus 4.61269734
MiMo-V2.51220880
GLM-51208190
GPT-5.51206045
Sonnet 51163518
GLM-5.3 Flash1140949
DeepSeek-V4-Pro1066426
GLM-5.21013158
GPT-5.41000803
MiniMax M3999506
Gemini 3.5 Flash987017
Gemini 3.6 Flash968123
GPT-6.1 Sol936267
Qwen3.6 Plus (2026-04-02)787999
GPT-5.6 Sol728001
GPT-5.6 Terra636144
Qwen3.7-Max (source unspecified)629583
Kimi K2.6511137
Kimi K2.5408822
Gemini 3 Flash Preview393736
GPT-6 Luna331093
Loading Atlas data…