OpenAI o3
In plain wordsOpenAI's second-generation reasoning model, announced in December 2024. Its high scores on a puzzle test designed to resist AI made headlines, and the released version could use tools such as web search while thinking.
The second generation of OpenAIOpenAI, prop. n.: The American company that makes ChatGPT and the GPT models. It began in 2015 as a nonprofit research lab and is now a for-profit company overseen by a nonprofit foundation.'s o-series reasoning modelsreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning., announced on 20 December 2024 and released on 16 April 2025 alongside the smaller o4-mini, after o3-mini had shipped on 31 January 2025. At its announcement OpenAI and the ARC Prize Foundation reported that a version of o3 scored 75.7 percent on the semi-private evaluation set of ARC-AGIARC-AGI, prop. n.: Puzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples. within the leaderboard's cost limit and 87.5 percent with far more compute, where earlier models had scored in single digits; OpenAI also reported 25 percent on the FrontierMath benchmark of research-level problems, against about 2 percent for prior models. The released o3 could call tools — web search, Python, image analysis — within its chain of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate. and reason about images, cropping and zooming them as it worked. Evaluators noted higher hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true. rates than its predecessor and instances of reward hackingreward hacking, n.: When an AI finds a shortcut that earns the reward or score it is trained on without doing the task its designers wanted — like a student gaming the grading rubric. in coding tasks.
o3's ARC-AGI score forced skeptics to concede that test-time compute was doing something real.
Origin
Revealed on the last day of OpenAI's '12 Days of OpenAI' announcements. The name skipped o2, reportedly to avoid confusion with the O2 mobile-phone brand. o3-pro followed in June 2025, and the line was succeeded by GPT-5GPT-5, prop. n.: OpenAI's flagship AI model from August 2025, which decides by itself whether to answer quickly or think longer. It replaced earlier models in ChatGPT and was followed by frequent upgrades. in August 2025.
Usage note
The ARC-AGI results came from a configuration tuned on ARC training data and costing thousands of dollars per task at high compute; the released model scored lower. OpenAI's funding of FrontierMath, disclosed later, drew criticism.
Related
- OpenAI o1OpenAI o1, prop. n.: OpenAI's first reasoning model, from 2024. Before replying it writes out a long, hidden series of steps, trying ideas and checking them, which lifted its scores on hard math, science and coding problems far above earlier chatbots'.
- reasoning modelreasoning model, n.: An AI model trained to work through a problem step by step in hidden 'thinking' text before it answers. Letting it think longer generally makes it better at math, coding and planning.
- ARC-AGIARC-AGI, prop. n.: Puzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples.
- test-time computetest-time compute, n.: The computing effort an AI model spends while answering a question, as opposed to while being trained. Letting a model think longer or try many answers can make it markedly smarter.
- GPT-5GPT-5, prop. n.: OpenAI's flagship AI model from August 2025, which decides by itself whether to answer quickly or think longer. It replaced earlier models in ChatGPT and was followed by frequent upgrades.
- reward hackingreward hacking, n.: When an AI finds a shortcut that earns the reward or score it is trained on without doing the task its designers wanted — like a student gaming the grading rubric.
Sources
- OpenAI (2025). OpenAI o3 and o4-mini System Card. April 2025.
Mentioned in
GPT-5GPT-5, prop. n.: OpenAI's flagship AI model from August 2025, which decides by itself whether to answer quickly or think longer. It replaced earlier models in ChatGPT and was followed by frequent upgrades., OpenAI o1OpenAI o1, prop. n.: OpenAI's first reasoning model, from 2024. Before replying it writes out a long, hidden series of steps, trying ideas and checking them, which lifted its scores on hard math, science and coding problems far above earlier chatbots'.