Tests explained · Test of real work

Graduate-level science questions

Biology, physics and chemistry questions written so they can't be answered by searching the web.

Run by

NYU and others

How to read it

The share of tasks the model got right. The thin line, where shown, is the likely range.

Used for

5 tasks on this site.

How the models did

Share of tasks done correctly · Higher is better

GPT-6 Astra and Gemini 3.8 Flash are neck and neck at the top, followed by Gemini 3.1 Pro and Gemini 3.6 Flash.

GPT-6 Astra · OpenAI96%

Gemini 3.8 Flash · Google DeepMind95%

Gemini 3.1 Pro · Google DeepMind94%

Gemini 3.6 Flash · Google DeepMind94%

GPT-5.6 Sol · OpenAI93%

Kimi K3 · Moonshot AI93%

DeepSeek V4 Pro · DeepSeek92%

GPT-5.6 Luna · OpenAI92%

Claude Sonnet 5 · Anthropic91%

Gemini 3.5 Flash-Lite · Google DeepMind83%

Claude Haiku 4.5 · Anthropic60%

Claude Fable 5.1 · AnthropicNot tested yet

Claude Opus 5.5 · AnthropicNot tested yet

DeepSeek V4.1 Flash · DeepSeekNot tested yet

GPT-6 Luna · OpenAINot tested yet

GPT-6 Sol · OpenAINot tested yet

Source: NYU and others via Epoch AI · CC BY 4.0

Where we use it

Each task page labels this test by how closely it matches the task.

Task	How close
Best AI for studying and homework	Tests a related skill
Best AI for medical and nursing students	Tests a related skill
Best AI for academic research	Tests a related skill
Best AI for maths and science	Tests this task
Best AI for chemistry	Tests this task