OpenAI Launches o3-mini: Cheaper Reasoning, Free for First Time

OpenAI released o3-mini on Friday, January 31, 2025, the newest model in its reasoning series, describing it as the most cost-efficient option in that series and making it available to free ChatGPT users for the first time. The company had previewed it in December alongside a more capable system, o3. The official announcement does not attribute the release to any particular competitor.

The issuer presents the release as wider access, but its system card makes access distinct from assurance: it documents risks involving disallowed content, jailbreaks, and hallucinations, and rates overall pre-mitigation risk as medium, including persuasion, CBRN, and model autonomy. Wider availability does not establish that every answer is correct or safe.

What a reasoning model is, and why it matters

Reasoning models such as o3-mini use additional computation before producing an answer. That is not the same as fact-checking, and it does not guarantee a correct result: the system card itself evaluates hallucination risk. The actual trade-off is among effort, latency, and measured performance on particular tasks, not between an unchecked answer and one certified as true.

O3-mini is optimized for STEM problems, specifically programming, math, and science. The official model page confirms that it does not accept image input. OpenAI compares it with the o1 family under particular effort settings and evaluations, rather than claiming universal superiority.

According to the company's published evaluation, expert testers preferred o3-mini responses to o1-mini 56% of the time and observed 39% fewer major errors on difficult questions. In its A/B tests, OpenAI measured responses that were 24% faster on average: 7.7 seconds versus 10.16. These are results under the manufacturer's protocol, not an accuracy rate for every query.

Reasoning to order: low, medium, and high

The most practical new feature in o3-mini is that it lets you dial in how much it "thinks." Through the API, developers can choose between three reasoning effort levels — low, medium, and high — so the model works harder or less hard depending on the task and tolerance for latency.

That flexibility translates into a range of performance relative to the previous family. "With low reasoning effort, o3-mini achieves comparable performance with o1-mini, while with medium effort, o3-mini achieves comparable performance with o1," OpenAI writes. "O3-mini with medium reasoning effort matches o1's performance in math, coding and science while delivering faster responses. Meanwhile, with high reasoning effort, o3-mini outperforms both o1-mini and o1."

In ChatGPT, the model is set to medium reasoning effort by default, which the company says provides "a balanced trade-off between speed and accuracy." Paid users can select "o3-mini-high" in the model picker for "higher intelligence" in exchange for slower responses. Whichever version users choose, the model can work with web search to find up-to-date answers with links to relevant sources, though OpenAI cautions that the feature is still a "prototype."

Free in ChatGPT, with tier-based limits

The other headline is access. O3-mini rolled out to all ChatGPT users this Friday, including free users, who can activate it with the new "Reason" button in the chat bar or by asking ChatGPT to "regenerate" a response. It's the first time OpenAI has offered a reasoning model at no cost.

Limits vary by subscription tier:

• ChatGPT Plus and Team: up to 150 queries per day.
• ChatGPT Pro: unlimited access.
• ChatGPT Enterprise and Edu: rolling out in about a week.

On ChatGPT Gov, the company gave no indication.

Price: what the announcement allows us to compare

OpenAI described o3-mini as its most cost-efficient reasoning model and said it retained o1-mini's low cost. However, the preserved primary announcement does not include a historical table with prices, units and a date that would support the claim that it was 63% cheaper. Nor does it attribute the release to DeepSeek. “More cost-efficient” is the manufacturer's positioning; a price comparison requires two dated tariffs expressed in the same unit.

The fine print: not the most powerful model

OpenAI is unusually upfront about what o3-mini is not. It's not the company's most powerful model to date, nor does it beat R1 across every benchmark.

The manufacturer's charts show that results depend on the selected effort. On AIME 2024, a competition-mathematics test, the low setting is compared with o1-mini, medium with o1, and high outperforms both on that evaluation. On GPQA Diamond, which uses doctoral-level science questions, OpenAI says high effort delivers performance comparable to o1, not that it exceeds it.

The harness matters too. For SWE-bench Verified, OpenAI declares a fixed subset of 477 tasks validated on its infrastructure. The system card separates 39% with the open Agentless scaffold from 61% with internal tools intended to elicit maximum capability. That is a 22-point difference without changing the model name. On FrontierMath, the company calls its figures provisional and distinguishes runs using Python from a chart without tools or a calculator. The announcement itself points to the system card as the source of truth for SWE-bench. A score describes the model, tools, sample and execution rules—not the model in isolation.

The takeaway is that o3-mini's pitch isn't about smashing absolute records, but about the cost-latency-reliability equation: answering well, fast, and cheap. "While o1 remains our broader general-knowledge reasoning model, o3-mini provides a specialized alternative for technical domains requiring precision and speed," OpenAI wrote.

Safety and availability

OpenAI says o3-mini surpasses GPT-4o on its challenging safety and jailbreak evaluations and attributes part of the result to deliberative alignment: training the model to reason over human-written safety specifications. The system card qualifies the scope: these are the manufacturer's evaluations and mitigations, and the model reached medium pre-mitigation risk in several categories. It was the first model OpenAI placed at medium risk for model autonomy, yet it still performed poorly on real-world machine-learning research evaluations relevant to self-improvement, which a high rating requires. This is not a safety certification for every use.

On the API, the model rolled out that Friday to selected developers in usage tiers 3–5 through Chat Completions, Assistants and Batch. It supported function calling, Structured Outputs, developer messages and streaming, but not image input. “Available in the API” therefore described a particular set of accounts, interfaces and modalities, not universal access.

What changes from here

The move has two concrete implications. For end users, putting free reasoning in ChatGPT lowers the barrier to a capability that was previously reserved for paying customers. For businesses and developers, adjustable effort allows them to calibrate spending: use the low tier for routine tasks and save the high tier for problems that warrant it.

OpenAI frames o3-mini as "another step in OpenAI's mission to push the boundaries of cost-effective intelligence," without attributing its timing to a rival. The transferable lesson is to separate the release's three layers: capability measured under a protocol, access constrained by plan and modality, and cost that can only be compared through dated prices in the same unit.