ChatGPT o3-mini is the new, more cost-effective version of the AI model

Contents

Shortly after the AI model DeepSeek from China stirred up the market, OpenAI is now launching ChatGPT o3-mini. The new AI model is designed to cost less money and technical resources and is tailored to scientific tasks in the STEM (science, technology, engineering and mathematics) sector.

The model has proven superior to its predecessor o1-mini in tests, OpenAI explains in its presentation. In tests, external experts preferred o3-mini to o1-mini in 56 percent of cases. The testers also found 39 percent fewer serious errors, according to the report. In the coding benchmarks, o3-mini achieved a hit rate of up to 49.3 percent and responded 49.3 percent faster on average.

o3-mini performance can be adjusted in stages

A new feature compared to o1-mini is the option to adjust the "reasoning efficiency" in stages. Users can choose the level of performance with which o3-mini should apply its ability to imitate logical thinking in order to answer the user's query. Users can choose between low, medium and high. Depending on whether a more complex problem requires higher performance or whether speed and latency are more important.

An internet search function has also been built into o3-mini. This allows the AI to provide answers to user queries with links to relevant internet sources. However, this function is still intended as an early prototype.

o3-mini partially outperforms o1

o3-mini also performs well in various AI benchmarks. At low reasoning effort, it achieves a performance comparable to OpenAI o1-mini, while at medium effort it achieves a performance comparable to o1. At the "high" level, o3-mini outperforms both o1-mini and o1.

The model should be able to solve complex software engineering tasks independently and thus achieves a success rate of 49.3 percent in the "SWE-bench Verified" benchmark, for example, which is the best performance achieved by an AI model from OpenAI to date.

For biology, chemistry and physics questions at doctoral level, o3-mini achieves a performance well above o1-mini with low reasoning effort. With high effort, o3-mini achieves a performance comparable to o1.

If you use o3-mini as an AI developer via an API, you get off cheaply: at 1.10 US dollars per million tokens, the model is comparatively inexpensive. The portal "The Decoder" draws a direct comparison with the o1 model, one of OpenAI's flagship AIs: here, the o3-mini token price is even 93 percent cheaper. The Decoder sees this as a reaction to the competitive prices of the Chinese AI DeepSeek. o3-Mini is initially only available as an API for certain developers in usage levels 3 to 5.

Also available free of charge

Access to o3-mini is also available via the free version of ChatGPT. There, the model now replaces o1-mini. The limit of 50 messages per day has been tripled to 150 messages per day. However, free o3-mini users can only select the "medium" performance level.

Users of ChatGPT Plus, Team, and Pro can use o3-mini from now on, while users of the Enterprise plan for companies should gain access in the course of February 2025.

(nen)