OpenAI has introduced a new AI model called ChatGPT o3-mini. This model is designed to be less expensive and require fewer technical resources. It is specifically tailored for tasks in the STEM fields, which include Mathematics, Computer Science, Natural Sciences, and Technology.
Compared to its predecessor, o1-mini, the o3-mini model has shown superior performance in tests. External experts preferred o3-mini over o1-mini in 56% of the cases. Additionally, testers noted 39% fewer severe errors with o3-mini. In coding benchmarks, o3-mini achieved an accuracy rate of up to 49.3% and responded, on average, 49.3% faster.
A new feature of o3-mini is the ability to adjust its “Reasoning Effort” in stages. Users can choose how much effort the model should use to mimic logical thinking when responding to queries. Options include low, medium, and high, depending on whether the task requires more power or if speed and latency are more important.
Another addition to o3-mini is a function for internet searches. This feature allows the AI to provide answers with links to relevant online sources, although it is still considered an early prototype.
In various AI benchmarks, o3-mini performs well. With low reasoning effort, it matches the performance of OpenAI’s o1-mini. At medium effort, it performs comparably to o1. On the “high” setting, o3-mini surpasses both o1-mini and o1.
The model can independently solve complex software engineering tasks and achieves a success rate of 49.3% in the “SWE-bench Verified” benchmark, marking the best performance by an OpenAI model so far.
For questions in biology, chemistry, and physics at the doctoral level, o3-mini performs significantly better than o1-mini with low reasoning effort. At high effort, it matches the performance of o1.
Developers using o3-mini via an API find it cost-effective. At $1.10 per million tokens, the model is relatively cheap. Compared to the o1 model, o3-mini’s token rate is 93% cheaper. This pricing is seen as a response to competitive prices from the Chinese AI DeepSeek. Initially, the o3-mini API is available to specific developers in usage levels 3 to 5.
ChatGPT’s free version also provides access to o3-mini, replacing o1-mini. The daily message limit has been increased from 50 to 150. However, free users can only choose the “medium” performance level.
Users of ChatGPT Plus, Team, and Pro can now use o3-mini. Enterprise plan users will gain access by February 2025.
