Atlas

Benchmarks

← All benchmarks

SuperGPQA

General QA · 2025-02-20

SuperGPQA is a 26,529-question multiple-choice benchmark of graduate-level knowledge and reasoning spanning 285 disciplines, built with expert annotators and an LLM-assisted filtering loop. Questions carry about ten options (9.67 on average), so uniform guessing earns roughly 10.3 percent. Scores report overall accuracy; shot count and chain-of-thought use are recorded on each score row.

Top models (higher is better)

ModelScore
Gemini 3 Pro Preview74.0
Qwen3.6-Max-Preview73.9
Qwen3.7-Max (source unspecified)73.6
Opus 4.672.5
Qwen3.6 Plus (2026-04-02)71.6
Qwen3.7-Plus (source unspecified)71.4
Kimi K2.671.3
Opus 4.570.6
Qwen3.5 397B A17B70.4
Qwen3.5 Plus (2026-02-15)70.4
DeepSeek-V4-Pro69.9
Kimi K2.569.2
GLM-5.168.0
GPT-5.267.9
Qwen3 Max Thinking (2026-01-23)67.3
Qwen3.5 122B A10B67.1
GLM-566.8
Qwen3.5-Omni-Plus66.4
Qwen3.6 27B66.0
Gemma 4 31B IT65.7
Qwen3.5-27B65.6
Qwen3-Max (2025-09-23)65.1
Qwen3-235B-A22B64.9
Qwen3 235B A22B Thinking 250764.9
Qwen3.6 35B A3B64.7
Qwen3 VL 235B A22B Thinking64.3
Qwen3.5 35B A3B63.4
Qwen3 235B A22B Instruct 250762.6
Gemini 2.5 Pro62.3
DeepSeek-R1-052861.7
Gemma 4 26B A4B IT61.4
Qwen3-Next-80B-A3B-Thinking60.8
Qwen3-VL-235B-A22B-Instruct60.4
DeepSeek-V3.159.8
Qwen3 VL 32B Thinking59.0
Qwen3 Next 80B A3B Instruct58.8
GPT-5 Mini58.6
Qwen3.5-9B58.2
Gemini 2.5 Flash57.8
Qwen3-Coder-Next57.5
Loading Atlas data…