Earlier rumors said that OpenAI would release multiple models, but on Monday of this week, there was only GPT-4.1 seriesAnd it was only available via API, so those looking forward to other models were surely wondering when they’d finally get to see them. OpenAI didn’t keep everyone waiting long—they’ve just officially launched their most powerful models yet, o3 and o4-mini, which are also the first OpenAI models capable of incorporating image-based reasoning. ChatGPT Plus and Pro users can use them right away, while free users get access to o4-mini.
OpenAI officially unveils o3 and o4-mini models with integrated image reasoning
OpenAI officially launches o3, o4-mini series models, besides being the most powerful currently, there are a few points that differ from previous models:
- Integrate visual thinking
- Can use all ChatGPT tools, including searching the web, using Python to analyze uploaded files and other data, doing in-depth reasoning on visual input, and even generating images.
Of course, it’s also stronger than previous OpenAI models in terms of performance. First, let’s look at the o3 model.
Compared to the previous generation o1 model, o3 has become stronger in multiple areas such as programming, mathematics, science, and visual perception. OpenAI stated that o3 set new records in benchmarks such as Codeforces, SWE-bench, and MMMU.
o3 is very well-suited for multifaceted analysis and complex queries, and it performs particularly well on visual tasks, with a major error rate about 20% lower than o1.
Meanwhile, the o4-mini is an optimized small reasoning model that balances speed and cost-effectiveness, serving as a replacement for the previous o3-mini. It performs quite well on math, programming, and vision tasks, and it is the highest-scoring model in the AIME 2024 and 2025 benchmarks.
o4-mini has higher usage limits, making it a great option for tasks that require multiple inference queries.
Below are the test data for each item.
In AIME 2024, 2025, and Codeforces tests, o4-mini clearly leads o3-mini:
For physics PhD-level knowledge and expert-level question accuracy across multiple fields, o3 and o4-mini achieved quite good results, also surpassing the previous generation by a significant margin:
In multimodal testing, o4-mini already scores significantly higher than o1, and o3 goes without saying:
The program editing accuracy comparisons are all between the high versions of each model:
In terms of instruction comprehension and tool operation performance, o3 can be said to perform quite impressively—just Python combined with web browsing is already able to achieve results comparable to Deep Research:
Regarding integrated image thinking, in simple terms, o3 and o4-mini don’t just “see” images—they also incorporate them into their reasoning. This opens up a全新的 problem-solving approach that融合s visual and textual reasoning.
What’s even more noteworthy is that OpenAI also mentioned these models can manipulate images in real time—like rotating, scaling, or deforming them. This suggests that in the future, ChatGPT might gradually take over many of the things image editing software can do.
Starting immediately, ChatGPT Plus, Pro, and Team users will see o3, o4-mini, and o4-mini-high in the model menu, replacing the original o1, o3-mini, and o3-mini-high:
ChatGPT Enterprise and Edu users are expected to be able to use it in a week.
Free users can enable “Reasoning” mode before asking to experience o4‑mini.
OpenAI also announced that it will launch o3-pro with full tool support within a few weeks. Pro users can continue to use o1-pro for now.
Do o3 and o4-mini have usage limits?
OpenAI stated that the usage limits for o3 and o4-mini across all plans remain the same as the previous version, meaning:
- ChatGPT Plus, Team, and Enterprise users can use the o3 model for 50 messages per week. They can use o4-mini for a total of 150 messages per day, and o4-mini-high for 50 messages per day.
- ChatGPT Pro’s o3, o4-mini, and 4o models are nearly unlimited.
- The free version of ChatGPT cannot use o3, and o4-mini has usage limits. The exact number has not been officially announced.
Source: KOCPC Chinese