— Lisan al Gaib (@scaling01) March 2, 2026
Update 02/03/26 – 10:35 am (IST): GPT-5.4 has now appeared in a second OpenAI Codex pull request. This time the code includes a direct function call using gpt-5.4 as the model argument, plus a slash command explicitly described as “toggle Fast mode for GPT-5.4.” X user @scaling01, who flagged the original PR last week, confirmed this is the second such mention.
Separately, @nicdunz claims alpha-gpt-5.4 was spotted in a public /models endpoint and teases an imminent rollout with major spec upgrades. There’s no independent sourcing to back that up, so treat it accordingly. Two PRs in one week referencing GPT-5.4 by name is notable on its own.
Original article published on February 28, 2026, follows:
An OpenAI engineer opened a pull request in the company’s public Codex GitHub repo yesterday, working on better image support for the view_image tool. The idea was to skip compression and preserve full-resolution images when a feature flag is on, and the model is new enough. The version cutoff in the original draft? “GPT-5.4 or newer.”
That one line got screenshotted immediately. Lisan al Gaib was among the first to post it on X.
This development came right after users Emilz, and am.will shared screenshots of GPT-5.4 showing up in the Codex app model selector. That too was accidentally leaked by an OpenAI employee, Tibo, on X. The post is now deleted.
The PR was quietly updated hours later to read “gpt-5.3-codex or newer,” with force pushes right after. Here’s a screenshot of it now:
GPT-5.3-Codex launched February 5 with faster speeds and better terminal and computer-use performance. Three weeks in, and a newer version is already being referenced in code.
The feature itself adds support for passing original PNG, JPEG, and WebP bytes straight to the Responses API with “detail”: “original” — clearly built with model-specific behavior in mind.
We spotted A/B tests recently that pointed to a ChatGPT 5.3 rollout with hints of a Thursday launch. That never came through. So take release timelines loosely.
This could very well be a slip-up, a copy-paste mistake, or a placeholder. Happens. But between the code wording and the model selector sightings in the Codex app, GPT-5.4 is clearly real enough to already be referenced internally. OpenAI ships fast, and this is just more proof of that.
Featured image credit: @LLMJunky / X


