Atlas

Benchmarks

← All benchmarks

KellyBench (2023/24 season, mean ROI)

Forecasting · 2026-04-09

General Reasoning's KellyBench test scenario: an agent bets through the 2023/24 English Premier League season (120 matchdays) in a sandbox, building prediction models from historical match and player data and staking from its bankroll on home, draw, away and over/under 2.5 goals at real bookmaker closing odds, at least one bet per matchday. The score is the mean return on the starting bankroll over five seeds, in percent; stakes come from the current bankroll, so -100 (ruin) is the floor and the return is otherwise unbounded. Rewards are deterministic changes in log-wealth with no LLM grader.

Top models (higher is better)

ModelScore
Opus 4.7-3.7
GPT-5.4-7.9
GPT-5.5-9.6
Opus 4.6-11.2
GLM-5.2-29.6
DeepSeek-V4-Pro-46.4
GLM-5-51.6
Kimi K2.6-60.4
Gemini 3.1 Pro Preview-66.0
Mistral Medium 3.5-77.9
Kimi K2.5-89.6
Mistral Large 3 675B Instruct 2512-98.8
Loading Atlas data…