OpenAI Announces "o1," a Reasoning Model That "Thinks Before It Answers"
On September 12, 2024, OpenAI announced a new line of models that take time to think before responding: “o1-preview,” and a lighter-weight version, “o1-mini.” Unlike the previous GPT series, which answers instantly, this design works through complex problems step by step, achieving major accuracy gains on hard problems in math, science, and coding. OpenAI positioned it as “a new paradigm.”
Details
- Design philosophy: Adopts a mechanism where the model builds up an internal chain of reasoning before answering, introducing a new approach that prioritizes accuracy over instant responses
- Benchmark results: Achieved 83% accuracy on problems at the level of the International Mathematical Olympiad (versus 13% for GPT-4o). In competitive programming, it ranked in the top 11% (89th percentile)
- Strength in specialized fields: OpenAI said the model achieved PhD-level performance on hard problems in physics, chemistry, and biology
- o1-mini: A lighter-weight version aimed at tasks that require reasoning ability, such as coding, but not broad world knowledge. It costs 80% less than o1-preview
- Availability: Available to ChatGPT Plus and Team users from launch day, with Enterprise and Edu access following the next week. Usage limits were initially fairly strict, at 30-50 messages per week
What happened next
The o1-preview approach later became the official “o1” model, releasing generally on December 5, 2024. From then on, OpenAI’s model development moved toward unifying “instant-response” and “deliberate reasoning” systems. The success of this reasoning model also spurred competitors like Google and Anthropic to release their own reasoning-focused models.