Model Drop

o3-mini

January 31, 2025 2 min read

o3-mini, which OpenAI shipped on January 31, 2025, is a small, low-cost reasoning model tuned for math, science and coding, and the first reasoning model offered to free ChatGPT users.

Model: o3-mini (o3-mini)
Maker: OpenAI
Ship date: January 31, 2025
Pricing: $1.10 / $4.40 per M tokens
Context: 200K tokens
Where you can use it:
- ChatGPT, free users via Reason
- The API for usage tiers 3 to 5 at launch
When to reach for it:
- Cheap, fast STEM and coding work
- High effort for the hardest problems
- Skip it for images; it has no vision
How it compares:
- Replaces o1-mini in ChatGPT's model picker
- Closed rival to the open DeepSeek R1
What makes it different:
- Low, medium and high reasoning effort
- First small reasoner with function calling
Early reactions: Simon Willison called it massively cheaper than o1 and said it generally benchmarks higher than GPT-4o and o1, but not across everything.
Read more: OpenAI o3-mini, OpenAI o3-mini System Card

The benchmarks

Scores are for o3-mini at high reasoning effort against OpenAI's own o1, o1-mini and o1-preview; o1's AIME and GPQA bars include 64-sample majority voting while o3-mini's are single tries, SWE-bench uses 477 of the 500 tasks, and no chart or table in the launch post shows o3-mini (high) trailing.

OpenAI's numbers

AIME 2024 (competition math): 87.3%. +4.0 over o1 (83.3%), +23.7 over o1-mini (63.6%).
GPQA Diamond (PhD-level science): 79.7%. +1.7 over o1 (78.0%), +19.7 over o1-mini (60.0%).
SWE-bench Verified (real GitHub fixes): 49.3%. +0.4 over o1 (48.9%), +8.0 over o1-preview (41.3%).
FrontierMath (research-level math, no tools): 9.2%. +3.7 over o1 (5.5%), +3.4 over o1-mini (5.8%).
LiveBench Coding (fresh coding problems): 84.6%. +17.2 over o1 (67.4%).
MMLU (broad knowledge test): 86.9%. +1.7 over o1-mini (85.2%).
SimpleQA (short factual questions): 13.8%. +6.2 over o1-mini (7.6%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 11.0. −0.4 under DeepSeek R1 (11.4), +1.2 over o1-mini (9.8), +1.6 over Gemini 2.0 Flash Thinking (9.4).
ARC-AGI-1 (abstract visual puzzles): 34.5%. +20.5 over o1-mini (14.0%), +18.7 over DeepSeek R1 (15.8%).
Aider Polyglot (multi-language code edits): 60.4%. +3.5 over DeepSeek R1 (56.9%), −1.3 under o1 (61.7%).