GPT-5.3-Codex

What it is

GPT-5.3-Codex is a coding model that OpenAI announced and released on 5 February 2026^[1]. OpenAI described it as its "first model that was instrumental in creating itself"^[1], and pitched it for work beyond writing code: searching repositories, running terminal commands and debugging^[1][2].

Where it came from

The model followed GPT-5.2-Codex and GPT-5.1-Codex-Max in OpenAI's Codex line^[6]. According to OpenAI, it combined the frontier coding performance of GPT-5.2-Codex with the reasoning ability and professional knowledge of GPT-5.2, and it could be steered mid-task without losing context^[7]. That last point matters in practice: a long coding session involves many corrections and changes of direction, and a model that keeps its place in the task can accept a new instruction partway through instead of starting over.

The model was co-designed for, trained with and served on NVIDIA GB200 NVL72 systems^[7]. ZDNET reported that it was 25% faster than its predecessor^[3].

Reported results

OpenAI said the model reached a new industry high on both SWE-Bench Pro and Terminal-Bench 2.0, and reported 77.3% on Terminal-Bench 2.0, 64.7% on OSWorld-Verified and 81.4% on SWE-Lancer IC Diamond^[7]. Benchmarks like these are useful mainly as a controlled comparison between models; they measure performance on prepared tasks rather than on the messy work of a real codebase.

SWE-Bench Pro differs from earlier software-engineering benchmarks in that it spans four programming languages instead of only Python, and OpenAI describes it as more contamination-resistant than SWE-bench Verified^[7]. Contamination is a recurring worry for coding benchmarks, because test problems that have been published online can end up in training data.

OpenAI also said that early versions of the model were used during development to debug its own training, manage its own deployment and diagnose its own test results and evaluations^[7]. That is the concrete basis for the "instrumental in creating itself" framing: the model was not only the product of the work, but a tool used inside it.

Availability and specifications

At launch the model was available in the Codex app and on the web, initially to ChatGPT Pro subscribers, who pay $200 a month; API access was planned^[1].

The API documentation lists a 400,000-token context window, of which up to 272,000 tokens can be input, up to 128,000 output tokens, a knowledge cutoff of 31 August 2025, and both text and image input^[4]. The context window is the amount of text the model can take in and produce at once, and input and output are billed separately. Pricing was set at $1.75 per million input tokens and $14 per million output tokens^[4].

The Spark variant

A smaller, text-only version, GPT-5.3-Codex-Spark, followed as a research preview on 12 February 2026^[1][6]. It ran on hardware from Cerebras, and InfoQ reported that it was about 15 times faster^[1][6]. A research preview is typically a limited release meant to gather feedback before wider availability.

Place in the line

GPT-5.4, released in March 2026, became the next model for Codex^[6]. OpenAI deprecated gpt-5.3-codex on 1 October 2026, with shutdown set for 1 April 2027^[5]. The model's life as an API offering was therefore short, which is now normal for coding models that are replaced within months.