STEM Reasoning
GPT-5.4
OpenAI
Claude Opus 4.8
Anthropic
Gemini 3 Pro
Google DeepMind
DeepSeek V4
DeepSeek
Grok 4
xAI
Qwen3 Max
Alibaba
Claude Sonnet 4.6
Anthropic
Llama 4.1 405B
Meta
Scores are illustrative sample data.
What it measures
STEM Reasoning targets the hard middle of scientific problem-solving: long derivations, counterfactuals, and multi-step tasks where the final answer hides a chain of reasoning. We grade both the answer and the steps.
Methodology
Expert-authored items — Specialists write from genuine difficulty, not retrieval, and hand over a verifiable reference solution with each item.
Step-level rubrics — A per-item rubric credits the reasoning chain, so sound-but-partial work counts — not just the final token.
Difficulty tiers — Stratifying and cross-checking each item keeps the distribution controlled and the score interpretable.
← Back to Leaderboard