OpenAI engineers have referenced GPT-5.4 in at least two separate pull requests to the company's public Codex GitHub repository, both of which were quickly modified or scrubbed after the community noticed. The first appeared on February 27, when an engineer working on image support for the view_image tool set a version cutoff at "GPT-5.4 or newer." The second surfaced today, containing a direct function call using gpt-5.4 as the model argument and a slash command described as "toggle Fast mode for GPT-5.4."
What actually leaked
The original pull request was mundane: adding support for passing uncompressed PNG, JPEG, and WebP bytes to the Responses API using a new "detail": "original" parameter. The interesting part was a conditional check gating the feature behind GPT-5.4 specifically, which suggests the model handles image input differently than its predecessors. The engineer later force-pushed to change the reference to "gpt-5.3-codex or newer," the kind of hasty edit that only draws more attention.
Around the same time, users Emilz and am.will shared screenshots of GPT-5.4 appearing in the Codex app's model selector. An OpenAI employee named Tibo apparently posted about it on X before deleting the post. X user @scaling01, who flagged the original PR, confirmed the second mention today.
The spec rumors are a different story
Separately, an account called @nicdunz claims to have spotted "alpha-gpt-5.4" in a public /models API endpoint and teased major specification upgrades. Some of the numbers circulating on social media include a 2-million-token context window, cross-session persistent state, and a redesigned memory architecture aimed at autonomous agents. None of this has independent sourcing. The Codex PRs confirm the model exists internally; they say nothing about what it can do.
Worth remembering where things stand: GPT-5.3-Codex launched on February 5 with faster speeds and improved terminal and computer-use performance. But the general-purpose GPT-5.3 still hasn't shipped to ChatGPT or the broader API. GPT-5.2 remains the current flagship, released on December 11 under what media outlets described as a "Code Red" response to Google's Gemini 3.
So OpenAI is already referencing the version after a model that hasn't publicly launched yet. That's either a sign of how fast the team ships or a reminder that internal version numbers don't always map to release timelines in the way outsiders expect.
Can OpenAI actually ship this fast?
The company's release cadence since GPT-5's August 7 launch has been aggressive: GPT-5.1, GPT-5.2, and GPT-5.3-Codex all arrived within about six months. But prediction markets aren't buying a quick GPT-5.4 rollout. On Manifold, traders have the probability of a March release in the low double digits, noting that OpenAI has never shipped two .X increments in a single month.
"The Codex PR leaks are real," one forecaster wrote. "But GPT-5.3 general-purpose hasn't shipped yet. Leak confirms the model exists internally; it doesn't confirm imminent release."
The context window claims are particularly hard to evaluate. GPT-5.2 offers 400,000 tokens through the API, already trailing Gemini's 1-million-token window. Jumping to 2 million would leapfrog Google, but the only source for that number is an unverified social media post. OpenAI's own release notes make no mention of it.
What this probably means
Two pull requests in a week referencing a model by name is more than a typo. GPT-5.4 exists inside OpenAI's infrastructure, at minimum as a development target. The image-handling changes in the PR suggest architectural differences from GPT-5.3, not just tuning improvements. But the spec sheet floating around social media, with its 2-million-token context and persistent agent memory, remains entirely unverified and should be treated that way.
OpenAI has not commented. The general-purpose GPT-5.3 rollout, whenever it arrives, will be the next concrete data point.
