Story Notes
GPT-5.4: The Tease, The Leaks, and OpenAI's Sprint
GPT-5.4: The Tease, The Leaks, and OpenAI's Desperate Sprint
OpenAI Buried Their Own News
Yesterday OpenAI shipped a genuinely good update. GPT-5.3 Instant made ChatGPT faster, more direct, and way less preachy. If you have ever asked it a simple question and gotten a three paragraph lecture about structural considerations before it answered, that is what they fixed. It was a real improvement and the kind of thing users had been asking for. OpenAI posted the announcement and waited for the coverage to roll in.
Nobody cared.
About an hour after that announcement went live, the official OpenAI X account posted five words that made everyone forget GPT-5.3 existed.
"5.4 sooner than you Think."
No blog. No details. No context. Just that. And the internet lost its mind.
What made it even wilder was that OpenAI did not even need to tease it. The community had already found it on their own.
The Leaks They Couldn't Scrub Fast Enough
In the week before that tweet, GPT-5.4 leaked three separate times out of OpenAI's own public Codex repository. Not from some anonymous source. Not from a rumor. From their own code.
The first slip happened on February 27. A pull request went up with code comments referencing "GPT-5.4 or newer" tied to a new full-resolution image feature that bypassed the usual compression. Someone noticed. OpenAI scrubbed it with a force-push. Then scrubbed it again. Seven times total before it was gone.
A few days later a second pull request appeared adding a "/fast" slash command. The description said plainly "toggle Fast mode for GPT-5.4." Also scrubbed.
Then Tibo, a verified OpenAI engineer at @thsottiaux, posted a status update saying Codex was back online after a brief outage. Innocent enough. Except a user named Corey Noles replied saying the error message had shown "something very interesting." Tibo's response was three words: "You saw nothing."
What Corey had seen was the live error message at the bottom of the screen. It read: "This user's access to gpt-5.4-ab-arm1-1020-1p-codexswic-ev3 has been temporarily limited." That is not a rumor. That is not a pull request someone dug up. That is the model name sitting in a live production error message for anyone to read. You do not make that up.
What's Real and What's Hype
The leaks confirmed three things: the model name GPT-5.4 is real and already in production systems, a full-resolution vision pipeline is coming with no more compressed thumbnails, and there is a new Fast mode priority tier.
Everything else initially came from the community running with those leaks and extrapolating. The thread that kicked it all off was posted by @rohanpaul_ai on March 2nd and racked up almost 100,000 views. He laid out the leaked pull requests, connected them to OpenAI's push toward agentic AI, and started speculating about what a model built for that future would need. Dozens of outlets ran with it as if it were confirmed.
But as of this morning there is harder reporting on top of all that speculation. The Information, which has solid sourcing inside OpenAI, just dropped the most specific details we have seen yet on what GPT-5.4 actually looks like.
According to that reporting, GPT-5.4 ships with a 1 million token context window, bringing it to parity with Gemini and Claude on long-context tasks. There is a new Extreme reasoning mode that lets the model spend significantly more compute on tough questions. It is built to run tasks that take hours, not minutes, with improved memory across multi-step workflows and lower error rates on complex jobs. The whole thing is designed around agents and automation, with Codex called out specifically.
And perhaps the most telling detail: this is being framed internally as part of OpenAI's shift to monthly model updates.
That is not just a product note. That is OpenAI telling the industry their release cadence has permanently changed.
The Markets Are Taking This Seriously
Three accidental leaks plus an official tease in five days. By the time the tweet dropped the story had already escaped and real money started moving.
On Manifold the main GPT-5.4 market jumped to 55% chance of release before April and 74% before June. There is even a "By March 5?" market that was trading between 40 and 55% this morning. That is people betting this drops in days, not months. Polymarket's broader next frontier model contracts shifted their curves forward overnight. When traders with actual skin in the game move that fast, the signal is usually real.
The Street Fight
The competitive pressure has been building for months. Anthropic keeps closing the gap on reasoning benchmarks. DeepSeek came out of nowhere and started shipping faster and cheaper than anyone expected. The old OpenAI playbook of dropping a flagship model once a year and riding it for twelve months stopped working. The market rewards whoever ships next, not whoever shipped best. OpenAI cannot afford to go quiet right now. Every week without a headline is a week someone else owns the conversation.
And that problem got a lot worse last week.
The Hole They Need to Dig Out Of
One week ago OpenAI was sitting on a genuine mess.
Anthropic had just refused to give the Pentagon unrestricted access to Claude over two concerns: fully autonomous weapons and mass domestic surveillance of American citizens. Anthropic CEO Dario Amodei said he could not in good conscience sign the deal. Defense Secretary Pete Hegseth responded by designating Anthropic a supply chain risk to national security, meaning any contractor doing business with the US military would be barred from using Claude.
Hours after Anthropic walked away, Sam Altman announced OpenAI had signed the deal.
The backlash was immediate. Claude shot to number one on the App Store. A movement called QuitGPT claimed 1.5 million people had cancelled subscriptions or taken action. Altman himself later admitted the whole thing had been rushed and that it "looked opportunistic and sloppy."
So when you see "5.4 sooner than you think" posted the day after Altman is doing damage control on the Pentagon deal, it is worth asking what this announcement is actually about. Maybe it is just the next model on the roadmap. Or maybe getting everyone focused on what is coming next is exactly what OpenAI needs right now.
The whole industry is watching to find out exactly how soon "sooner than you think" actually is.
Update: March 5, 2026 - 12:35 AM
Remember Tibo? The verified OpenAI engineer who told Corey Noles "you saw nothing" after the live error message exposed the model name?
He just posted this.
"I feel a disturbance in the force and yet May the fourth is a long time from now... interesting."
98,000 views in a few hours. The replies are reading it one way: May 4th is a long time from now, meaning this does not drop in May. It drops much sooner. Possibly today.
The same engineer who accidentally confirmed the model name is now teasing the release window. Whether that is a coincidence or not is left as an exercise for the reader.
