Overview
On January 31, 2025, OpenAI released o3-mini — a lightweight reasoning model priced for broad API consumption while maintaining strong performance on STEM benchmarks.
The key differentiator from o3 and from competitors like DeepSeek R1: o3-mini had full web browsing access, making it the first reasoning model in its class that could ground its answers in current information.
Pricing
- Input: ~$0.55/million tokens
- Output: ~$4.40/million tokens
This was roughly 60% cheaper than o1-mini and dramatically cheaper than full o3 (which could cost $10-100+ per complex query depending on reasoning depth).
Performance
o3-mini performed comparably to o1 on most STEM benchmarks:
- AIME 2024: ≈85% (vs o1 ≈83%)
- GPQA Diamond: competitive with o1
- Codeforces: solid performance, though not Claude 3.7’s level
The key advantage over DeepSeek R1: web browsing + function calling + OpenAI’s API infrastructure.
Significance
o3-mini marked the moment reasoning models became commoditized as a developer tool, not just a research showcase. The combination of strong reasoning + web access + low price made it the default choice for AI coding assistants and agentic applications through 2025.