OpenAI has introduced its latest generative AI model family, code-named Strawberry, officially referred to as OpenAI o1. This release marks a significant development, with two versions available initially: o1-preview and o1-mini. While the models can be accessed via ChatGPT Plus and Team subscribers, enterprise and educational users will gain early access next week. Despite the advanced features of the new model, including its ability to reason through tasks holistically, o1 is currently expensive and limited in functionality.
The o1 model stands out for its self-fact-checking ability, made possible by its chain of reasoning feature. OpenAI claims that o1 can perform more complex tasks by taking additional time to analyze a query. This capability allows the model to handle tasks requiring multi-step reasoning, such as legal analysis or product strategy development. According to OpenAI, the longer o1 processes a problem, the more accurate its responses become, a feature enhanced by the reinforcement learning approach used during its training.
In terms of pricing, o1-preview costs $15 per 1 million input tokens and $60 per 1 million output tokens, making it significantly more expensive than OpenAI’s earlier GPT-4o model. Additionally, while o1 has been praised for its reasoning capabilities, it is rate-limited, with weekly message caps of 30 for o1-preview and 50 for o1-mini. Although the model includes image analysis features, they have been temporarily disabled pending further testing.
Feedback from early testers has been positive, with some noting that o1 excels at tasks like legal brief analysis and logic-based challenges. For example, in a test related to the International Mathematical Olympiad, o1 outperformed GPT-4o by solving 83% of problems compared to GPT-4o’s 13%. However, despite its strengths, users have reported occasional issues, such as hallucinations and slow response times for complex queries, with some responses taking up to 10 seconds.
OpenAI is not alone in exploring advanced reasoning for AI models. Google DeepMind recently published research showing that extending compute time during task processing can significantly improve model performance. This competition highlights the growing field of AI reasoning development, with OpenAI noting that it chose not to display o1’s raw chains of thought to maintain a competitive edge.
Although OpenAI leads with the o1 release, the company will face challenges in making the model more widely accessible and reducing costs. Future upgrades could see o1 reasoning over extended periods — hours, days, or even weeks — to further enhance its problem-solving abilities.
Featured Image courtesy of DonanimHaber
Follow us for more updates on GPT-o1.