Best AI models for science

Ranked by GPQA Diamond: very hard multiple-choice questions in biology, physics and chemistry written by PhD experts, where random guessing scores 25%. Tested independently by Epoch AI.

Updated September 30, 2026

Our picks

#ModelGPQA DiamondiOverall indexiInputiOutputiContexti
1
95.8%
166.6$10$501.1M
2
95.6%
—$2$101M
3
Gemini 3.8 FlashGoogle DeepMind
95.4%
157.1$0.75$3.751M
4
Gemini 3.7 FlashGoogle DeepMind
94.8%
157.7$0.75$3.751M
5
94.6%
159.1$30$1801.1M
6
Gemini 3.1 Pro PreviewGoogle DeepMind
94.4%
154.9$2$121M
7
Gemini 3.6 FlashGoogle DeepMind
94.1%
154.4$0.75$3.751M
8
94.0%
156.5$2$6500K
9
Claude Opus 5Anthropic
93.9%
162.7$5$251M
10
93.5%
162.0$2$101.1M

per 1M tokens · USD

Quick answers

Which AI model is best for science questions?
GPT-6 Astra from OpenAI leads this ranking (GPQA Diamond: 95.8%). It costs $10 per million input tokens and $50 per million output tokens.
Is there a cheaper alternative?
Gemini 3.8 Flash from Google DeepMind (GPQA Diamond: 95.4%) costs $0.75 / $3.75 per million tokens, about 13× less than GPT-6 Astra.

How we rank

Scores come from tests run independently by Epoch AI (CC BY 4.0), not from the companies’ own announcements. Prices come from OpenRouter’s public API. Both are refreshed automatically several times a day, so this ranking follows new releases and price changes.

Other rankings

CodingMathCheapFreeOpen sourceLong contextImages & PDFsRankings →