OpenAI introduces o1 reasoning models
A new series learns to spend more time reasoning before answering.
OpenAIA new reasoning approach
OpenAI introduced o1-preview and o1-mini on September 12, 2024, opening a new model series focused on difficult reasoning tasks. Plus and Team subscribers received access in ChatGPT, while API access initially required a qualifying usage tier. These were early releases rather than replacements for every GPT-4o use. OpenAI
The important change was how additional computation could improve an answer. OpenAI reported that reinforcement learning taught o1 to work through problems, revise approaches and correct mistakes. Performance improved both with more training and with more computation spent reasoning before a response. That made time spent solving a task an explicit part of the capability story. OpenAI
What the preview delivered
The research report distinguished the released preview from a stronger o1 model. On its single-sample AIME mathematics evaluation, OpenAI reported 44.6% accuracy for o1-preview, 9.3% for GPT-4o and 74.4% for o1. Results using many candidate answers were reported separately. The stronger model and extra-computation results should not be treated as the performance every preview user received. OpenAI
The preview also lacked familiar features, including web browsing and file or image uploads; its API initially omitted function calling and streaming. OpenAI said GPT-4o remained more capable for many common tasks. Users saw a model-generated reasoning summary rather than the raw chain of thought, limiting what they could directly inspect about the process behind an answer. OpenAI OpenAI
Sources & attribution
- Organizational disclosure 12 Sept 2024o1 preview release announcement
OpenAI. Original developer announcement.
- Organizational disclosure 12 Sept 2024Learning to reason with LLMs
OpenAI. Original developer research report.