A new OpenAI machine learning mode has appeared. On Friday, the company released o3-mini, and now it is available to try. Moreover, for the first time, OpenAI is making one of its “reasoning” models available to free ChatGPT users. If you want to try it yourself, click the Reasoning button below the message composer to get started.
According to OpenAI, o3-mini is faster and more accurate than its predecessor, o1-mini. During A/B testing, the company found that o3-mini was 24% faster than o1. Moreover, given “average” reasoning effort, the new model can approach the performance of the more expensive o1 system in some math, coding, and science tests. Like other OpenAI reasoning models, o3-mini will show you how it arrived at the answer, rather than just answering the prompt. Notably, the model works with ChatGPT Search out of the box, allowing it to scour the web for the latest information and useful links. OpenAI says it is working to integrate search into all of its reasoning models.
“The release of OpenAI o3-mini marks another step in OpenAI’s mission to push the boundaries of cost-effective intelligence. By optimizing reasoning for STEM domains while keeping costs low, we are making high-quality AI even more accessible,” OpenAI said. “This model continues our success in lowering the cost of intelligence – since the launch of GPT-4, the price per token has dropped by 95% – while maintaining top-notch reasoning capabilities. As AI adoption expands, we remain committed to leading the way by creating models that balance intelligence, efficiency, and security at scale.”
With today’s announcement, the o3-mini will replace the o1-mini in the model selection. Additionally, OpenAI is tripling the message limit for Plus and Team ChatGPT users from 50 messages per day on the o1-mini to 150 messages per day on the o3-mini. OpenAI’s newly launched Pro-tier, costing $200 per month, offers unlimited access to the new system.
When OpenAI first unveiled o3 and o3-mini late last year, CEO Sam Altman said the latter would be available “around the end of January.” Altman gave a more specific timeline on January 17, when he wrote on X that OpenAI “plans to release a version in a couple of weeks.”
Now that it’s here, it’s safe to say that o3-mini is arriving with a sense of urgency. On January 20, the same day that Altman attended Donald Trump’s inauguration, the Chinese company DeepSeek quietly released its R1 thought chain model. By January 27, the company’s chatbot had surpassed ChatGPT as the most downloaded free app in the US App Store after going viral. DeepSeek’s overnight success wiped out $1 trillion in market value and almost certainly left OpenAI by surprise.
Last week, OpenAI announced that it was working with Microsoft to identify two accounts that the company claims may have been distilling its models. Distillation is the process of transferring the knowledge of an advanced artificial intelligence system to a smaller, more efficient system. Distillation is not a controversial practice. DeepSeek has used distillation on its own R1 model to train its smaller algorithms; in fact, OpenAI’s terms of service allow distillation as long as users do not train competing models on the company’s AI results.
OpenAI does not name DeepSeek directly. “We know that companies based in [China] and beyond are constantly trying to outperform the models of the leading American AI companies,” an OpenAI spokesperson recently told The Guardian. However, David Sachs, President Trump’s artificial intelligence advisor, was more outspoken, claiming that there is “strong evidence” that DeepSeek “distilled knowledge from OpenAI’s models.”
