The o3-mini preview was released in December 2024, along with the o3 model. OpenAI has been criticized by some as ceding the AI race to Chinese companies like DeepSeek.

ChatGPT developers tout the new model as “powerful” and “affordable.” A spokesperson says the o3-mini marks an important step in making advanced AI accessible to more people.

openai cnet
OpenAI faces considerable pressure from Chinese competitors. Photo: Cnet

Unlike other large language models, an inference model like o3-mini validates itself before delivering results to users. This helps avoid some of the pitfalls of conventional models. Inference models will provide slower but more reliable – albeit imperfect – solutions in domains like physics.

The o3-mini is tuned to solve STEM problems ( science , math, programming). OpenAI claims it is on par with the o1 and o1-mini in terms of performance, but is faster and cheaper.

Independent programmers rated the o3-mini's responses as better than the o1-mini's. Additionally, on real-world tricky questions, the o3-mini made 39% fewer “major errors” than the o1-mini, and provided “clearer” responses 24% faster.

o3-mini is available to all ChatGPT users from January 31, but ChatGPT Plus and Team users can ask more, up to 150 queries per day. ChatGPT Pro users get unlimited use. o3-mini will be available on ChatGPT Enterprise and ChatGPT Edu within a week.

Paid ChatGPT users can select o3-mini from the drop-down menu on the app, while free users click on the new “Reason” button in the chat box.

The cost of o3-mini is $0.55/1 million input tokens and $4.4/1 million output tokens, where 1 million tokens correspond to about 750,000 words. It is 63% cheaper than o1-mini and also competitive with DeepSeek R1. DeepSeek is charging $0.14/1 million input tokens and $2.19/1 million output tokens for R1.

In ChatGPT, the o-3 mini is set to medium inference mode to balance response speed and accuracy. Premium users can choose higher mode, which provides better quality but slower response times. Regardless of the version used, the model looks for the latest replies and related links.

(According to TechCrunch)