Atlas

Benchmarks

← All benchmarks

Concept Synth ABD v1.2 - Full

General QA · 2026-02-21

The fully observed, closed-world split of ABD v1.2. A synthesized abnormality rule is valid when the repaired theory is satisfiable in every prompt world.

Top models (higher is better)

ModelScore
Opus 4.6100.0
Gemini 3.1 Pro Preview99.5
DeepSeek-V3.296.1
Grok 4.1 Fast95.5
Grok 487.2
GPT-5.284.9
Gemini 3 Pro Preview77.1
GPT-5.476.4
Kimi K2 Thinking60.0
GPT-4o8.7
Loading Atlas data…