OpenAI has launched o3-mini, its latest and most cost-efficient model in the reasoning series. Available in both ChatGPT and the API, o3-mini builds upon the strengths of its predecessors while introducing key advancements. This model excels in STEM fields, particularly science, math, and coding, offering exceptional performance at a reduced cost and latency.
o3-mini is the first small reasoning model from OpenAI to support highly requested developer features. These include function calling, structured outputs, and developer messages, making it immediately production-ready. Like previous models, o3-mini supports streaming and offers three reasoning effort options (low, medium, and high) for optimized performance depending on the task’s complexity and urgency. While o3-mini doesn’t handle visual reasoning (use o1 for that), it represents a significant leap forward in text-based reasoning capabilities.
Access to o3-mini is tiered. ChatGPT Plus, Team, and Pro users gain immediate access, with Enterprise access following in February. o3-mini replaces o1-mini in the model picker, offering higher rate limits (tripled for Plus/Team users to 150 messages/day) and lower latency. It also integrates search functionality for up-to-date answers with web source links. Free plan users can also experience o3-mini by selecting “Reason” or regenerating a response, marking the first time a reasoning model is available to free users.
o3-mini provides a specialized alternative to o1, which remains OpenAI’s broader general knowledge reasoning model. In ChatGPT, o3-mini defaults to medium reasoning effort for a balance of speed and accuracy. Paid users can also select o3-mini-high for more complex tasks, with Pro users having unlimited access to both versions.
o3-mini shines in STEM reasoning. At medium effort, it matches o1’s performance in math, coding, and science, but with faster responses. Expert testing reveals o3-mini produces more accurate and clearer answers with stronger reasoning. Testers preferred o3-mini 56% of the time over o1-mini and saw a 39% reduction in major errors on difficult questions. It even matches o1’s performance on challenging evaluations like AIME and GPQA at medium effort, and outperforms it at high effort. Performance benchmarks across various STEM domains like competition math, PhD-level science questions, research-level mathematics, competition coding, and software engineering showcase o3-mini’s prowess.
Speed and efficiency are also key features. o3-mini is faster than o1, with a 24% faster response time in testing. This translates to an average response time of 7.7 seconds compared to o1-mini’s 10.16 seconds. The time to the first token is also significantly reduced.
Safety is paramount. o3-mini was trained using deliberative alignment, reasoning about safety specifications before responding. It surpasses GPT-4 on safety and jailbreak evaluations. Thorough safety assessments, external red-teaming, and evaluations were conducted before deployment.
OpenAI’s o3-mini represents a significant advancement in cost-effective AI. By optimizing STEM reasoning and lowering costs (95% reduction in per-token pricing since GPT-4), OpenAI makes high-quality AI more accessible. The company remains committed to pushing the boundaries of intelligent, efficient, and safe AI at scale.