OpenAI releases o3-pro, a souped-up version of its o3 AI reasoning model

OpenAI has introduced o3-pro, its latest and most advanced AI reasoning model to date. This new offering builds on the original o3 model, which was launched earlier this year. Unlike traditional AI systems that may approach problems in a linear fashion, reasoning models like o3-pro tackle challenges methodically, making them particularly effective in complex fields such as mathematics, physics, and programming. Starting Tuesday, o3-pro is available for ChatGPT Pro and Team users, replacing the previous o1-pro model. OpenAI has announced that Enterprise and Educational users will gain access to this enhanced version the following week. Additionally, o3-pro has been integrated into OpenAI’s developer API as of this afternoon, with a pricing structure set at $20 per million input tokens and $80 per million output tokens. For context, a million input tokens roughly equate to around 750,000 words, slightly longer than the classic novel "War and Peace." In a recent changelog, OpenAI highlighted that expert evaluations show a strong preference for o3-pro over o3 across all tested categories, particularly in vital areas such as science, education, programming, business, and writing support. Reviewers noted that o3-pro consistently outperformed in clarity, comprehensiveness, adherence to instructions, and overall accuracy. The capabilities of o3-pro are enhanced by its access to a suite of tools, which enable web searches, file analysis, reasoning with visual data, Python programming, and personalized responses through memory integration. However, users should note that responses from o3-pro may take longer compared to its predecessor, o1-pro. Currently, there are some limitations with o3-pro. Temporary chats are disabled while OpenAI addresses a technical issue, and the model is unable to generate images. Additionally, it does not support Canvas, OpenAI’s AI-powered workspace feature. On a positive note, internal testing reveals that o3-pro excels in various AI benchmarks, outperforming Google’s leading AI model, Gemini 2.5 Pro, in math skills as measured by AIME 2024. It also surpasses Anthropic’s Claude 4 Opus in the GPQA Diamond test, which assesses PhD-level science knowledge.