Atlas

Benchmarks

← All benchmarks

SciPredict

Forecasting · 2026-01-15

Forecast accuracy on 405 real scientific experiment outcomes in physics, biology, and chemistry under the No Background Knowledge (NBK) condition. Published results average three trials and their error bars are run standard deviations, not 95 percent confidence intervals. Observation sampling metadata preserves the repeated mean without multiplying the distinct-task count.

Top models (higher is better)

ModelScore
Gemini 3 Pro Preview25.3
Opus 4.523.1
Sonnet 4.522.6
Opus 4.122.2
Gemini 3 Flash Preview22.2
GPT-5.220.6
o3-mini19.8
DeepSeek-V319.2
Llama 3.3 70B Instruct18.2
o317.9
Gemini 2.5 Pro17.0
Qwen3 32B17.0
Qwen3-235B-A22B16.6
O4 Mini16.2
Llama 3.1 8B Instruct14.7
Loading Atlas data…