Skip to content

DeepSeek

DeepSeek-R1

Open-weight reasoning model under the MIT licence. DeepSeek said it performed on par with OpenAI's o1 (self-reported).

Announced
Access
Public

Benchmark scores

TestScoreBest on the trackerHow it was measured
GPQA DiamondPhD-level science questionsPhD experts: 70%Saturated72%Gemini 4 Argon 96%Independentdefault settings (DeepSeek-R1)Run 26 May 2025
ARC-AGI-2visual puzzles, model aloneAverage person: 60%Saturated1.3%GPT-6 Astra 95%IndependentARC Prize config "Deepseek R1" (R1)
OTIS Mock AIMEcompetition mathsSaturated53%GPT-5.5 100%Independentdefault settings (DeepSeek-R1)Run 26 Feb 2025

Compare every model on the same tests →

Sources