OpenAI released o1 series AI models with enhanced reasoning

OpenAI has unveiled the o1 series of AI models—o1-preview and o1-mini—designed to enhance reasoning in complex tasks across science, coding, and mathematics. These models allocate more computational time to "think" before responding, significantly improving performance in advanced problem-solving.

Key Features:

1. Advanced Reasoning: Trained with reinforcement learning focused on complex reasoning, the o1 models excel in solving challenging problems that previous models struggled with.
2. Benchmark Achievements:
   1. Science Disciplines: Performance comparable to PhD students in physics, chemistry, and biology.
   2. International Mathematics Olympiad: Achieved 83% accuracy, a substantial increase from GPT-4o's 13%.
   3. Coding Competitions: Ranked in the 89th percentile on Codeforces, indicating strong coding proficiency.

Enhanced Safety Measures:

The o1 models incorporate a new safety training approach, utilizing their reasoning abilities to adhere strictly to safety and alignment guidelines. They demonstrate increased resistance to "jailbreaking," with o1-preview scoring 84 out of 100 in safety tests, compared to GPT-4o's score of 22.

Model Variants:

1. o1-preview: The larger model with robust reasoning capabilities and extensive world knowledge.
2. o1-mini: A smaller, faster, and more cost-effective version, particularly effective in coding and math tasks, and 80% cheaper than o1-preview.

Availability:

Starting September 12, 2024, the o1 models are accessible to ChatGPT Plus and Team users, with plans to extend access to ChatGPT Enterprise and Edu users the following week. Developers on API usage tier 5 can begin prototyping with both models. Initial weekly rate limits are set at 30 messages for o1-preview and 50 for o1-mini, with plans to increase after further testing.