OpenAI has announced the release of the o1-preview, the first in a groundbreaking series of AI models designed to enhance reasoning capabilities. These models are engineered to spend more time deliberating before generating responses, enabling them to tackle complex tasks and solve challenging problems in fields such as science, coding, and mathematics.

Enhancing AI's Thought Process

The o1 series represents a significant advancement in artificial intelligence by simulating human-like reasoning. Through specialized training, these models learn to:

- Deliberate Thoroughly: Spend additional time analyzing problems before responding.

- Refine Strategies: Experiment with different approaches to find the most effective solutions.

- Self-Correct: Recognize and learn from mistakes to improve future performance.

In rigorous testing, the upcoming update to the o1 model has demonstrated performance comparable to that of PhD students on challenging benchmarks in physics, chemistry, and biology. Notably, the model has shown exceptional proficiency in mathematics and coding:

- Mathematics: Solved 83% of problems on a qualifying exam for the International Mathematics Olympiad, a substantial improvement over the 13% solved by the earlier GPT-4o model.

- Coding: Achieved the 89th percentile in Codeforces programming competitions.

For a detailed analysis, refer to OpenAI's [technical research post](#).

Prioritizing Safety and Alignment

With increased capabilities, the importance of safety measures becomes paramount. OpenAI has developed a new safety training methodology that leverages the o1 models' advanced reasoning to ensure adherence to strict safety and alignment guidelines. By understanding and applying these rules within context, the models demonstrate stronger compliance, even under challenging conditions.

In "jailbreaking" tests—designed to assess a model's resistance to attempts at bypassing safety protocols—the o1-preview model scored 84 out of 100, a significant improvement over GPT-4o's score of 22. This indicates a robust ability to maintain safety standards.

OpenAI's comprehensive safety approach includes:

- Rigorous Testing: Utilizing a Preparedness Framework for thorough evaluations.

- Advanced Red Teaming: Engaging in extensive testing to identify and mitigate potential risks.

- Governance Oversight: Implementing board-level review processes with a Safety & Security Committee.

- Global Collaboration: Partnering with U.S. and U.K. AI Safety Institutes to share research and provide early access for testing.

More details can be found in the [system card](#) and the [research post](#).

Applications and Beneficiaries

The o1 models are poised to benefit professionals and researchers dealing with complex problems, including:

- Healthcare Researchers: Assisting in annotating intricate cell sequencing data.

- Physicists: Generating advanced mathematical formulas necessary for quantum optics.

- Software Developers: Building and executing multi-step workflows across various industries.

Introducing o1-mini: Efficient and Cost-Effective Coding

Alongside the o1-preview, OpenAI is releasing **o1-mini**, a streamlined model optimized for coding tasks. Key features include:

- Enhanced Speed: Faster performance for code generation and debugging.

- Cost Efficiency: 80% more affordable than the o1-preview model.

- Targeted Application: Ideal for projects requiring strong reasoning capabilities without extensive world knowledge.

Accessing the o1 Models

For ChatGPT Users

- ChatGPT Plus and Team Users: Both o1-preview and o1-mini are accessible starting today via the model selector. Initial weekly limits are set at 30 messages for o1-preview and 50 for o1-mini. OpenAI plans to increase these limits and introduce automatic model selection based on user needs.

- ChatGPT Enterprise and Education Users: Access to both models will be available starting next week.

- ChatGPT Free Users: There are plans to extend o1-mini access to all free users in the near future.

For Developers

- API Access: Developers with API usage at tier 5 can begin experimenting with both models today, with a rate limit of 20 requests per minute. The API currently does not support function calling, streaming, or system messages.

Future Developments

This release marks an early preview of what the o1 series will offer. OpenAI is committed to:

- Ongoing Improvements: Regular updates to enhance model capabilities.

- Feature Expansion: Adding functionalities such as browsing, file and image uploading to increase utility.

- Parallel Progress: Continuing advancements in both the GPT series and the new o1 series to provide a diverse range of AI solutions.