OpenAI has released o3-mini, a new reasoning model that brings a more advanced type of AI response to people who use the free version of ChatGPT. The release matters because OpenAI’s reasoning models had previously been limited to paid Pro and Plus bundles, putting them outside the reach of most users.
The company says o3-mini is faster, cheaper, and more accurate than its predecessor. It arrives at a moment when reasoning models are becoming a central battleground in AI, with Microsoft rolling OpenAI's reasoning model o1 out to its Copilot users and DeepSeek drawing attention for lower-cost performance claims.
What o3-mini changes for ChatGPT users
The biggest shift is access. With o3-mini, people using the free version of ChatGPT can try one of OpenAI’s reasoning models for the first time. According to the source article, this marks the first time that the vast majority of people will have access to one of these models.
To use o3-mini, a ChatGPT user can select Reason when starting a new prompt. That option signals a different kind of model behavior from a standard quick-response chatbot experience.
Reasoning models are designed to work through difficult prompts step by step. The source describes this as a chain of thought technique, where the model processes a problem in stages, checks its own path, and can correct mistakes before it gives a final answer.
That can be useful for demanding tasks. The source points to complex problems, including PhD-level math problems, and advanced prompts that previous models may not have handled well. The basic idea is simple: when a question requires careful logic rather than a fast surface-level answer, a reasoning model may be the better tool.
Speed improves, but the pause remains part of the product
Reasoning models can deliver more thorough and accurate responses, but they often take longer to answer. That delay comes from the additional work happening before the response appears.
OpenAI claims that o3-mini responds 24% faster than o1-mini. That is an important improvement because wait time is one of the practical tradeoffs of using a model built to reason through a task. A faster reasoning model can make the experience feel less like a specialist tool and more like something people might use regularly.
Even with that speed gain, the model is still built around a different response pattern. Users should expect it to be most relevant when the prompt benefits from careful reasoning, not when the goal is a simple answer that any standard model can handle quickly.
The cost picture is better, not solved
OpenAI is also positioning o3-mini as a cheaper reasoning model than o1-mini. The source says o1-mini is 20 times more expensive to run than its equivalent non-reasoning model, GPT-4o mini. OpenAI says o3-mini costs 63% less than o1-mini per input token.
That reduction is meaningful, but it does not make reasoning models inexpensive compared with non-reasoning models. At $1.10 per million input tokens, o3-mini is still about seven times more expensive to run than GPT-4o mini.
This cost gap helps explain why free access is notable. Reasoning models require more computation to produce an answer, and the source notes that they are estimated to have much higher energy costs than other types because of that larger computation load.
The release also lands shortly after DeepSeek drew intense attention in the AI world. DeepSeek’s new model performs just as well as top OpenAI models, according to the source, while the Chinese company claims it cost roughly $6 million to train. That comparison is being made against the estimated cost of over $100 million for training OpenAI’s GPT-4, though the source also notes that a lot of people are interrogating DeepSeek’s claim.
DeepSeek’s reasoning model costs $0.55 per million input tokens, which is half the price of o3-mini. That puts pressure on OpenAI’s cost story: o3-mini is cheaper than o1-mini, but the broader market is pushing the question of how low reasoning model costs can go.
Safety becomes harder as reasoning gets stronger
The same capabilities that make reasoning models useful also create safety challenges. OpenAI trained its o-series models using a technique called deliberative alignment. In plain terms, the source says this means the models reference OpenAI’s internal policies at each step of reasoning so they do not ignore rules.
But OpenAI has found that o3-mini, like the o1 model, is significantly better than non-reasoning models at jailbreaking and at challenging safety evaluations. That matters because a model with stronger reasoning ability may also be harder to control.
o3-mini is the first model to receive a medium risk score on model autonomy. The rating is tied to the model being better than previous models at specific coding tasks, which OpenAI says indicates greater potential for self-improvement and AI research acceleration.
There is an important limit in the source: the model is still bad at real-world research. If it were better at that, OpenAI would rate it as high risk and restrict its release.
Why this release matters
o3-mini is not just another model update. It moves reasoning AI from a paid-only feature into the hands of free ChatGPT users, while also showing the competing priorities shaping AI development: better answers, faster responses, lower operating costs, and tighter safety controls.
The model’s release suggests reasoning systems are becoming more mainstream, but not frictionless. They can be more useful for complex prompts, yet they still carry higher costs than non-reasoning models. They can work more carefully, yet their stronger capabilities make safety evaluations more difficult.
For users, the practical takeaway is straightforward. When a prompt requires careful logic, advanced problem-solving, or a more deliberate answer, o3-mini gives free ChatGPT users a new option. For OpenAI, the release shows progress on access and cost, while also underscoring that reasoning models remain expensive and more challenging to govern.