OpenAI o3

OpenAI o3 is a reflective generative pre-trained transformer (GPT) model developed by OpenAI as a successor to OpenAI o1. It is designed to devote additional deliberation time when addressing questions that require step-by-step logical reasoning.^[1][2] On January 31, 2025, OpenAI released a smaller model, o3-mini,^[3] followed on April 16 by o3 and o4-mini.^[4]

The OpenAI o3 model was announced on December 20, 2024, with the designation "o3" chosen to avoid trademark conflict with the mobile carrier brand named O2.^[1] OpenAI invited safety and security researchers to apply for early access of these models until January 10, 2025.^[5] Similarly to o1, there are two different models: o3 and o3-mini.^[3]

On January 31, 2025, OpenAI released o3-mini to all ChatGPT users (including free-tier) and some API users. OpenAI describes o3-mini as a "specialized alternative" to o1 for "technical domains requiring precision and speed".^[6] o3-mini features three reasoning effort levels: low, medium and high. The free version uses medium. The variant using more compute is called o3-mini-high, and is available to paid subscribers.^[3][7] Subscribers to ChatGPT's Pro tier have unlimited access to both o3-mini and o3-mini-high.^[6]

On February 2, OpenAI launched OpenAI Deep Research, a ChatGPT service using a version of o3 that makes comprehensive reports within 5 to 30 minutes, based on web searches.^[8]

On February 6, in response to pressure from rivals like DeepSeek, OpenAI announced an update aimed at enhancing the transparency of the thought process in its o3-mini model.^[9]

On February 12, OpenAI further increased rate limits for o3-mini-high to 50 requests per day (from 50 requests per week) for ChatGPT Plus subscribers, and implemented file/image upload support.^[10]

On April 16, 2025, OpenAI released o3 and o4-mini, a successor of o3-mini.^[4]

Reinforcement learning was used to teach o3 to "think" before generating answers, using what OpenAI refers to as a "private chain of thought".^[11] This approach enables the model to plan ahead and reason through tasks, performing a series of intermediate reasoning steps to assist in solving the problem, at the cost of additional computing power and increased latency of responses.^[12]

o3 demonstrates significantly better performance than o1 on complex tasks, including coding, mathematics, and science.^[1] OpenAI reported that o3 achieved a score of 87.7% on the GPQA Diamond benchmark, which contains expert-level science questions not publicly available online.^[13]

On SWE-bench Verified, a software engineering benchmark assessing the ability to solve real GitHub issues, o3 scored 71.7%, compared to 48.9% for o1. On Codeforces, o3 reached an Elo score of 2727, whereas o1 scored 1891.^[13]

On the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) benchmark, which evaluates an AI's ability to handle new logical and skill acquisition problems, o3 attained three times the accuracy of o1.^[1][14]