OpenAI’s newest reasoning model o3-pro surpasses rivals on multiple benchmarks, but it’s not very fast – News

OpenAI is upping the ante in artificial intelligence reasoning with the launch of o3-pro today, claiming it’s the most advanced model of its kind.

The o3-pro model is a more powerful version of the original 03 reasoning model, which debuted earlier this year. Reasoning models are more intelligent than standard generative AI models like the GPT family, as they work through problems in a step-by-step way, carefully considering their responses to ensure they’re more accurate. They’re especially useful when it comes to solving problems in areas such as physics, coding and math.

According to OpenAI, o3-pro will think for even longer than the original 03 model, and provide “the most reliable responses.” It’s helped by the fact it can access various third-party software tools to complete assigned tasks, and it’s designed to appeal to developers and researchers who want the highest level of detail and accuracy in their AI outputs.

In its release notes detailing the new model, OpenAI said its evaluations show that reviewers “consistently prefer 03-pro over 03 in every tested category,” especially in domains such as science, programming, business and education. It adds that reviewers also rated it “consistently higher” in terms of its clarity, accuracy, instruction-following and comprehensiveness.

The company first launched o3 in April, and followed shortly thereafter with 04-mini, claiming that they can “think with images.” It said o3-pro is based on the same model as o3, but it has been tweaked to improve its performance by making it think for longer.

Reasoning models have become one of the hottest battlegrounds in the AI industry, and OpenAI is challenged here by the likes of Google LLC, Anthropic PBC and xAI Corp., as well as the Chinese startup DeepSeek Ltd., which have also developed thinking models.

OpenAI said o3-pro has been given access to more tools than its earlier reasoning models, so it can browse the web for information, analyze and reason visual inputs, use the Python coding language and personalize its responses. But one of the drawbacks of this is that it takes significantly more time to come up with answers.

The company admitted this, saying users will notice that its responses “typically take longer” than o1-pro. For this reason, it recommends the model for the most challenging questions “where reliability matters more than speed, and waiting a few minutes is worth the tradeoff.”

The responses will indeed likely take several minutes. Hyerbolic Labs Inc. co-founder and Chief Technology Officer Yuchen Jin posted on X that he had to wait more than three minutes for o3-pro to respond to the input “Hi, I’m Sam Altman”:

Slow responsiveness isn’t the only problem users may notice with o3-pro. OpenAI said temporary chats are currently disabled because of a “technical issue” with the model, and it can’t generate images either. Moreover, OpenAI’s AI workspace product Canvas does not support o3-pro.

However, the sheer power of o3-pro may make it worthwhile for some users. The model racked up some impressive benchmark results, surpassing Google’s best reasoning model, Gemini 2.5 pro on the AIME 2024 test that measures math proficiency. It also beat Anthropic’s top model Claude 4 Opus on the GPQA Diamond benchmark that evaluates PhD-level science skills.

OpenAI said o3-pro is available now for all ChatGPT Pro and Team users, where it replaces the older o1-pro, which was its previous-best reasoning model. ChatGPT Enterprise and Edu subscribers will get access to the model next week, and it can also be found in the company’s application programming interface for developers.

But users should beware that all of the extra thinking comes at a cost – o3-pro is priced at $2 per million input tokens and $80 per million output tokens. One million tokens is equivalent to around 750,000 words, which is approximately the length of Leo Tolstoy’s classic novel “War and Peace.” Input tokens refer to what the user inputs into the model, while output tokens are what the model spits out in response.

Perhaps in order to soften the blow, OpenAI also announced today that it’s slashing the cost of its o3 model by 80% for both inputs and outputs. The news came from OpenAI Chief Executive Sam Altman, who said on X that he’s doing this to encourage broader experimentation with the model.

Image: News/Microsoft Designer