OpenAI has released o3-pro, a new artificial intelligence model that the company claims is its most capable yet. The model is available for ChatGPT Pro and Team users, with Enterprise and Education users gaining access next week.
o3-pro is an enhanced version of OpenAI’s o3 reasoning model. Unlike conventional AI models that generate immediate responses, reasoning models work through problems step by step. This approach allows them to perform more reliably in complex domains like mathematics, physics, and computer programming. Learn more about Reasoning AI on our explainer article.
Key capabilities and limitations
The new model has access to various tools that expand its functionality. It can search the web, analyze uploaded files, interpret visual content, run Python code, and personalize responses using memory. OpenAI states that expert reviewers consistently preferred o3-pro over the standard o3 model across all tested categories, particularly in science, education, programming, business, and writing assistance.
However, these enhanced capabilities come with trade-offs. OpenAI acknowledges that o3-pro responses take significantly longer to complete compared to its predecessor, o1-pro. The company recommends using the model for challenging questions where accuracy matters more than speed.
Early users have reported extreme processing times for simple queries.
Pricing and technical constraints
o3-pro carries a premium price tag through OpenAI’s developer API. The model costs $20 per million input tokens and $80 per million output tokens. For context, one million input tokens equals approximately 750,000 words. This represents significantly higher costs compared to the standard o3 model, which is priced at $2 for input and $8 for output tokens.
The model currently has several limitations. OpenAI has temporarily disabled private chat sessions while resolving technical issues. o3-pro cannot generate images, requiring users to switch to other models like GPT-4o for image creation. The Canvas workspace feature is also not yet supported.
Performance benchmarks
According to OpenAI’s internal testing, o3-pro achieves strong results on industry-standard AI benchmarks. On AIME 2024, which tests mathematical abilities, the model outperforms Google’s Gemini 2.5 Pro and Anthropic’s Claude Opus 4. OpenAI uses a “4/4 reliability” evaluation system, where models must correctly answer questions in all four attempts to be considered successful.
The launch comes as reasoning models become increasingly competitive among AI companies. Google, Anthropic, xAI, and Chinese companies like DeepSeek have all released their own reasoning-focused models.
Despite early reports of high costs and slow processing times, some users have praised o3-pro’s performance. Former Apple Vision Pro designer Ben Hyak noted improvements in the model’s ability to understand its environment and choose appropriate tools for tasks.
Sources: OpenAI, VentureBeat, TechCrunch