OpenAI o3

What o3 is

o3 is a reasoning model from OpenAI, meaning a model trained to work through a problem internally before answering instead of producing the first plausible response. That internal deliberation costs extra computation, so answers come more slowly, but it helps on problems that need several steps of derivation. OpenAI announced o3 on 20 December 2024 and released it on 16 April 2025 together with the smaller o4-mini^[1].

Why it is called o3

The model is named o3 rather than o2 to avoid a trademark conflict with the mobile carrier O2^[1]. It continues OpenAI's o-series: the company's API documentation lists o3 as the successor of o1 and GPT-5 as the model that succeeded it^[2].

What it can do

OpenAI highlighted two capabilities at release. First, o3 had full access to the tools inside ChatGPT as well as custom tools supplied through function calling, and it was trained to reason about when and how to use them — deciding for itself whether a question calls for a search, a calculation or a code run^[7]. Second, o3 and o4-mini were the first OpenAI models able to integrate images directly into their chain of thought, so that a picture could become part of the reasoning process rather than only an input read at the start^[7].

Results

On GPQA Diamond, OpenAI reported 87.7%^[1]. On SWE-bench Verified, which tests whether a model can resolve real software issues in a code repository, o3 scored 71.7% against 48.9% for o1^[1]. On Codeforces, the competitive-programming platform, it reached an Elo rating of 2727, against 1891 for o1^[1]. On the ARC-AGI benchmark it reached three times o1's accuracy^[1]. OpenAI described o3 as its most powerful reasoning model and said it set a new state of the art on Codeforces, on SWE-bench without a custom model-specific scaffold, and on MMMU^[7]. In evaluations by external experts, o3 made 20 percent fewer major errors than o1 on difficult real-world tasks^[7]. At equal latency and cost to o1, OpenAI reported that o3 delivered higher performance inside ChatGPT^[7].

API specifications

Through the API, o3 had a context window of 200,000 tokens and could produce up to 100,000 output tokens, with a knowledge cutoff of 1 June 2024 and support for both text and image input. It was priced at $2 per million input tokens and $8 per million output tokens^[2].

Where else it was used

A version of o3 powered deep research, a ChatGPT capability launched on 2 February 2025^[1][6]. A fine-tuned version called codex-1 powered the Codex agent in May 2025^[1][5].

Place in the line, and retirement

OpenAI's GPT-5 system card names gpt-5-thinking as o3's successor^[4]. On 28 May 2026 OpenAI announced that o3 would leave ChatGPT on 26 August 2026, and the o3 API snapshot was deprecated on 11 June 2026, with shutdown set for 11 December 2026^[1][3]. In the o-series it therefore occupies the middle position: the generation that followed o1, widened the gap on coding and science benchmarks, and was in turn replaced by GPT-5^[2].