OpenAI o3-mini is a Large language model developed by OpenAI, released on January 31, 2025. It is a smaller variant of the OpenAI o3 model, designed as a cost-effective alternative for developers and users who need the reasoning capabilities of the larger model with lower compute requirements. o3-mini is built on a Generative AI Transformer (architecture) architecture, leveraging Deep learning techniques for step-by-step logical reasoning.
The model is part of the o3 series, which succeeded the OpenAI o1 model. o3-mini was introduced to provide a balance between performance and cost, targeting technical domains such as coding, mathematics, and science, where precision and speed are essential. It offers three reasoning effort levels - low, medium, and high - allowing users to trade off response latency and quality based on their needs.
History and Release
The OpenAI o3 model was first announced on December 20, 2024. The name "o3" was chosen over "o2" to avoid a trademark conflict with the mobile carrier brand O2. OpenAI invited safety and security researchers to apply for early access to the model until January 10, 2025. Following the announcement, two variants were planned: o3 and o3-mini.
On January 31, 2025, OpenAI released o3-mini to all ChatGPT users, including free-tier users, and to some API users. OpenAI described o3-mini as a "specialized alternative" to o1 for technical domains requiring precision and speed. The free version of ChatGPT uses the medium reasoning effort level, while paid subscribers can access o3-mini-high, a variant that uses more compute. ChatGPT Pro tier subscribers received unlimited access to both o3-mini and o3-mini-high.
Shortly after, on February 2, OpenAI launched OpenAI Deep Research, a service that uses a version of o3 to generate comprehensive reports within 5 to 30 minutes based on web searches. On February 6, in response to pressure from rivals such as DeepSeek R1, OpenAI announced updates to enhance the transparency of o3-mini's thought process. On February 12, OpenAI increased rate limits for o3-mini-high for ChatGPT Plus subscribers, raising them from 50 requests per week to 50 requests per day, and added file and image upload support.
On April 16, 2025, OpenAI released o3 and o4-mini, the successor to o3-mini. Following that, on June 10, OpenAI released o3-pro, which the company claimed was its most capable model yet. In a statement, OpenAI recommended using o3-pro for challenging questions where reliability matters more than speed, noting that waiting a few minutes is worth the tradeoff. On May 28, 2026, OpenAI announced that o3 would be retired from ChatGPT on August 26, 2026, following a 90-day sunset period, but this change did not affect the API.
Capabilities and Performance
OpenAI used Reinforcement learning to train o3-mini to "think" before generating answers, a method referred to as a "private chain of thought." This approach allows the model to plan ahead and reason through tasks, performing intermediate reasoning steps to solve problems, albeit at the cost of additional computing power and increased response latency.
o3-mini demonstrates significantly better performance than o1 on complex tasks, including coding, mathematics, and science. According to OpenAI, the full o3 model achieved a score of 87.7% on the GPQA Diamond benchmark, which contains expert-level science questions not publicly available online. On SWE-bench Verified, a software engineering benchmark, o3 scored 71.7% compared to o1's 48.9%. On Codeforces, o3 reached an Elo score of 2727, whereas o1 scored 1891. On the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) benchmark, o3 attained three times the accuracy of o1.
These benchmarks highlight the model's advanced reasoning abilities, making it suitable for tasks that require logical deduction and multi-step problem-solving. The introduction of o3-mini extends these capabilities to a broader audience by reducing operational costs.
Technical Design and Architecture
Like its predecessor, o3-mini is a Neural network based model that uses a Transformer (architecture) architecture. It incorporates Multi-Head Attention mechanisms and Positional Encoding to process input sequences effectively. The model is trained using Curriculum Learning and RLHF (Reinforcement Learning from Human Feedback), which helps align its outputs with human expectations.
One of the key features of o3-mini is its adjustable reasoning effort levels, which control the depth of the Chain-of-thought process. This feature allows developers to optimize for speed or accuracy depending on the application. The model also supports iterative refinement, enabling users to guide it towards more precise answers.
Pricing and Availability
OpenAI positioned o3-mini as a cost-effective solution for developers. The company noted that it is significantly cheaper than the full o3 model, with the price being 80% lower in some configurations. This pricing strategy aims to make advanced reasoning capabilities accessible to a wider range of applications, from Amazon Web Services and Microsoft Azure cloud deployments to mobile apps.
o3-mini is available through the OpenAI API and within the ChatGPT interface. As of its initial release, it was accessible to free-tier users with limited usage, while paid subscribers had higher rate limits belows mentioned.
Comparisons and Impact
In the competitive landscape of Artificial intelligence, o3-mini faces rivals such as Anthropic's Claude and Google DeepMind's models. The release of o3-mini was part of OpenAI's response to the emergence of cost-competitive models like DeepSeek R1, which pressured the company to offer more affordable reasoning options. The model's performance on benchmarks like ARC-AGI and Codeforces demonstrates its capability to handle complex reasoning tasks, making it a valuable tool for Machine learning researchers and developers.
The introduction of o3-mini has influenced the broader AI ecosystem, encouraging other companies to focus on efficiency and affordability. It has also sparked discussions about the transparency of reasoning processes, leading to updates in how OpenAI communicates the model's thought patterns.