OpenAI releases GPT-5.3-Codex

Background

Codex is the name OpenAI uses for its coding agent — software that works on code on a user's behalf. In OpenAI's own account, Codex began as an agent that writes and reviews code ^[4]. Days before the new model appeared, the company released a desktop app for Codex ^[3], adding another place, beyond the browser and the code editor, where the agent could be run.

The release

On 5 February 2026 OpenAI released GPT-5.3-Codex for its Codex coding agent ^[1]. The company described it as the most capable agentic coding model to date, and presented the launch as a widening of what the agent does at all: Codex, it said, had gone from an agent that writes and reviews code to one that can do nearly anything developers and professionals do on a computer ^[4].

The model was made available with paid ChatGPT plans everywhere Codex runs — in the app, the command-line interface, the IDE extension and on the web — with API access still to come ^[4].

Testing: two games, millions of tokens

OpenAI said that during testing it had the model build two games, iterating autonomously over millions of tokens on generic prompts such as "fix the bug" ^[4]. The detail is about how the work was driven: short, unspecific instructions, with the model deciding what to change next and carrying on for a long stretch of tokens — the units of text, in this case mostly code, that a model reads and writes — rather than a person directing each step.

OpenAI also stated that the model had been instrumental in creating itself ^[1][2]. The claim is about the part the model played in the work that produced it, and it was reported alongside the launch rather than demonstrated to outside observers in the materials at hand ^[2].

A faster variant

A faster research-preview variant, GPT-5.3-Codex-Spark, followed on 12 February 2026, a week after the main release ^[1]. A research preview is a build released for early use rather than as a finished product, and the variant was characterised chiefly by speed.

Use, and a security commitment

A month after the launch, Fortune reported 1.6 million weekly Codex users ^[3].

OpenAI also committed $10 million in API credits to cybersecurity defence, building on a $1 million grant programme it had started in 2023 ^[4]. API credits are prepaid allowances for using a model through OpenAI's programming interface; the commitment therefore amounts to funded model usage for defensive security work, such as reviewing code or hunting for vulnerabilities, rather than an equivalent sum of cash.

What the release changed

Three things stand out in OpenAI's framing. The first is scope: the agent's job description moved from writing and reviewing code to a much broader range of computer work ^[4]. The second is autonomy: the testing account has the model running for millions of tokens from prompts as thin as "fix the bug" ^[4]. The third is self-reference: a coding model that was, by the company's account, used in its own creation ^[1][2].

Set against the numbers reported a month later — 1.6 million weekly Codex users ^[3] — the release marked a step in the agent's move from a tool for reading and writing code towards a general worker on a machine ^[4].