OpenAI o1-preview release notes / what's new

OpenAI has introduced the o1 series, a new class of AI models designed to spend more time "thinking" before responding. This shift toward reasoning-centric architecture allows these models to solve significantly harder problems in science, coding, and mathematics than previous iterations.

Reasoning Capabilities and Performance

OpenAI o1 models are trained to refine their thinking process, recognize mistakes, and try different strategies before delivering a final answer. This approach results in substantial performance gains in technical domains:

  • Mathematics: In a qualifying exam for the International Mathematics Olympiad (IMO), the reasoning model scored 83%, compared to 13% for GPT-4o.
  • Coding: The models reached the 89th percentile in Codeforces competitions.
  • Science: In challenging benchmark tasks across physics, chemistry, and biology, the next model update performs similarly to PhD students.

For many common tasks, GPT-4o remains more capable in the near term because the o1-preview currently lacks features such as web browsing, file uploads, and image support.

Safety and Alignment Training

OpenAI implemented a new safety training approach that leverages the models' own reasoning capabilities to adhere to alignment guidelines. By reasoning about safety rules in context, the models can apply them more effectively.

Internal testing on "jailbreaking" (attempts to bypass safety rules) showed a significant improvement: the o1-preview model scored 84 on a scale of 0-100, while GPT-4o scored 22.

To support these capabilities, OpenAI has integrated the following governance measures:

  • Preparedness Framework: Rigorous testing and evaluations.
  • External Collaboration: Formalized agreements with the U.S. and U.K. AI Safety Institutes, including granting them early access to research versions of the model.
  • Internal Review: Board-level review processes conducted by the Safety & Security Committee.

Model Variants: o1-preview and o1-mini

OpenAI is releasing two versions of the reasoning series to balance performance and efficiency:

  • OpenAI o1-preview: The primary reasoning model designed for complex, multi-step problems.
  • OpenAI o1-mini: A faster, cheaper alternative specifically optimized for coding. It is 80% cheaper than o1-preview and is intended for applications that require reasoning without the need for broad world knowledge.

Availability and Access

Access to the o1 series is being rolled out across different tiers:

  • ChatGPT Plus and Team: Access to o1-preview and o1-mini via the model picker. As of September 17, 2024, rate limits are 50 queries per week for o1-preview and 50 queries per day for o1-mini.
  • ChatGPT Enterprise and Edu: Access began the week following the initial September 12 release.
  • Developers: API access is available for those in usage tier 5, with an initial rate limit of 20 RPM. The current API does not support function calling, streaming, or system messages.
  • Free Users: OpenAI plans to bring o1-mini access to all ChatGPT Free users.

Future Roadmap

OpenAI will continue to develop the GPT series alongside the o1 series. Planned improvements for the o1 models include the addition of web browsing, file uploading, and image uploading capabilities.

Sources