Atlas

Benchmarks

← All benchmarks

MatFlowBench Computational

Science · 2026-10-09

The 25 computational MatFlowBench tasks, in which agents plan and run density functional theory and related calculations (Quantum ESPRESSO, Wannier90, Phonopy) on electronic structure, topology, spin and anomalous Hall transport, magnetism, strain, Wannier models and lattice dynamics. The verifier reads the agent's calculation output files; the score is the share of passing rollouts, two per task.

Top models (higher is better)

ModelScore
Claude Opus 5.564.0
GPT-6 Astra58.0
Claude Sonnet 5.556.0
GPT-6.1 Sol46.0
Kimi K318.0
Grok 4.714.0
Hy4 Preview14.0
Muse Spark 1.38.0
Gemini 3.8 Flash2.0
Loading Atlas data…