The Open-Source AI Stack
RSS
All models

Models · Compare

OpenAI o3-mini vs DeepSeek-R1

Rows highlighted in warm gray are where the models differ. Numbers carry their as-of date and primary source.

Specs

Field A: OpenAI o3-mini B: DeepSeek-R1
Released 2025-01-31
Developer OpenAIDeepSeek
Openness ProprietaryOpen
License ProprietaryMIT
OSI-approved noyes
Data released nono
Training code nono
Architecture unknownmoe
Total params 671B
Active params 37B
Experts
Context window 200K128K
Attention unknownmla
Position enc. unknownrope-yarn
Pretraining tokens
Post-training rlhfsft, grpo, rejection-sampling
Training hardware H800
$/M input $1.10$1.35
$/M output $4.40$4.20
Output tok/sec 145.30

Benchmarks

Missing scores render as not reported; never inferred. Bold highlights the leader per benchmark.

General reasoning

MMLU 90.8 2025-01-22
MMLU-Pro 79.1 2026-05-21 84.9 2026-05-21
GPQA-Diamond 79.7 2025-01-31 71.5 2025-01-22

Code

SWE-Bench Verified 49.3 2025-01-31
LiveCodeBench 71.7 2026-05-21 77.0 2026-05-21

Math

MATH 97.3 2026-05-21 97.3 2025-01-22
AIME 2024 87.3 2025-01-31 79.8 2025-01-22
AIME 2025 76.0 2026-05-21

Context · A

Released January 31, 2025 as a smaller, faster o-series reasoning model with three selectable effort levels (low, medium, high). Positioned by OpenAI as a specialized alternative to o1 for technical domains requiring precision and speed at a fraction of o1's cost.

Context · B

The first openly-released reasoning model competitive with OpenAI o1. The R1-Zero variant demonstrated that pure-RL post- training without SFT could elicit chain-of-thought reasoning. MIT-licensed weights, distillations into 1.5B-70B sizes.

OpenAI o3-mini detail → · DeepSeek-R1 detail →