GPT 5.4 leaks (confirmed)

There's been tons of rumors and leaks about OpenAI's GPT-5.4.

As of today, we can confirm these are not just rumors.

Yesterday OpenAI shipped GPT-5.3 Instant, and their next model already leaked.

Not once. Not twice. Three times in one week.

From OpenAI's own GitHub code, their error logs, and an employee screenshot.

Then this morning The Information talked to their source at OpenAI and confirmed it with new details.

Before we get into 5.4, quick note on 5.3 Instant.

It's now the default ChatGPT model for all users.

The big focus is making ChatGPT less annoying to talk to.

27% fewer hallucinations. Better writing. Better search. Fewer unnecessary caveats and refusals.

And apparently it's been told to stop telling users to calm down.

Have you ever seen a situation where one person tells another person to calm down and it works as intended?

Where the other person goes "oh, you're right, let me just relax."

Never.

Calm down is basically fighting words.

And for some reason the previous GPT models were really over-indexing on that expression.

TechCrunch literally ran a headline that said ChatGPT's new model will stop telling you to calm down.

That's the official news.

The unofficial news is way more interesting.

How GPT-5.4 Leaked Three Times

First leak, February 27.

An OpenAI engineer opened a change in the public Codex GitHub repo adding full-resolution vision support.

The code set the minimum model version to 5.4.

A model that doesn't officially exist yet.

Screenshots went out on X immediately.

OpenAI's response was seven force pushes over five hours to change that number from 5.4 to 5.3.

Updating the codebase seven times over five hours is not done to fix a typo.

That's a fire drill.

Second leak, March 2.

Different engineer, different change.

They added a /fast command to Codex.

Codex is open source so anybody can see what they're changing in real time.

And the original code said "toggle fast mode for GPT-5.4."

Direct name reference.

Scrubbed within three hours.

Third leak, March 3.

An OpenAI employee posted a screenshot where GPT-5.4 was visible in the model selector dropdown.

Post was deleted, but screenshots are forever.

Separately, journalist Corey Noles hit a cybersecurity block where the error message referenced an internal model ID for GPT-5.4.

That's confirmation this model is running on OpenAI's servers right now.

What The Information Confirmed

Stephanie Palazzolo at The Information confirmed it from a person with direct knowledge.

That name keeps popping up in AI reporting. Well connected, has sources.

So this is bonafide.

1 million token context window.

Up from 400,000 in GPT-5.2.

The rumor was 2 million. It's 1 million.

Still 2.5x the previous context and puts OpenAI back on par with Google and Anthropic who already support 1 million tokens.

Extreme reasoning mode.

So you know how you have light reasoning, medium reasoning, high reasoning.

If you keep climbing apparently you get to extreme thinking.

This isn't just a bump up. It's significantly more time.

We're talking hours, not seconds.

You ask a question, leave it, come back hours later and it has your answer.

The closest comparison might be Gemini's deep research mode where sometimes it runs 40 minutes plus compiling information.

But this seems to be an extension of pure reasoning, not just tool calls.

Very curious to see what the right use cases and benchmarks will be for this.

Better long-running task performance.

This is a big one.

The model is better at remembering details across many steps.

If the user has certain requirements or many different steps, current models can forget what they were doing or lose some minor detail along the way.

You might have seen the AI researcher at Meta whose email agent just deleted all her messages because the context window got wiped and it forgot she'd said to confirm before deleting.

Things like that.

This new model should be better at not doing that.

Less prone to mistakes over long chains of reasoning.

Think of it as improving how reliable agents are over time instead of just raw intelligence.

Full resolution vision.

No more compressing your uploads.

If you're sending screenshots of code, medical images, diagrams, architectural drawings, schematics, the model will see every pixel.

Before, uploading an image would compress it and you'd lose data. That's going away.

Priority inference / fast mode.

A new service tier system with standard and fast.

If you're using AI agents for real-time tasks where latency matters, where they need to think fast and move fast, switching to that priority mode is going to be important.

And then there's the garlic question.

You might recall rumors from months ago about OpenAI's "Garlic" AI.

Their leaders said it fixed a lot of training issues and would help challenge Gemini.

We don't know if GPT-5.4 incorporates Garlic, but if it does, that could be the secret sauce behind these improvements.

The Deliberate Pace

You might have noticed how fast OpenAI is shipping.

We're seeing a new version roll out pretty much once a month now.

And OpenAI is confirming this is deliberate. Not an accident.

The reason makes sense.

Think about GPT-5. We built it up in our minds so much that when it actually came out, a lot of people saw it as a disappointment.

But if you look at the METR scaling chart, that exponentially growing curve, there's consistent, steady progress.

In fact it's beginning to look like it's accelerating.

So how could any one model on that curve be a letdown?

It's because of cadence.

If you have regular releases then you expect incremental improvements. Not too much hype, not too much letdown.

GPT-5 was awaited so long that by the time it arrived everyone expected something grand.

The reality was the train was right on schedule. It wasn't late, it wasn't early. It was just expectations.

The Numbers

OpenAI is at about 910 million weekly active users.

They were hoping to hit 1 billion by the end of 2025 so they're a little behind.

GPT-5.1 and 5.2 reinvigorated that growth after a slowdown following GPT-5.

But now they're taking a hit from the QuitGPT movement.

Over 2.5 million people have pledged to leave.

ChatGPT uninstalls surged 295%.

Claude is now ahead in first-time app downloads.

Anthropic just hit $19 billion in annual revenue, up from $14 billion just weeks ago.

And as of today, Anthropic is back in talks with the Pentagon about a deal.

I mentioned a few videos back that there was a chance this might happen.

A lot of people in different departments wanted to keep using Claude.

Fingers crossed it works out.

I'm preparing a full dive into all of this because there's a lot happening and I want to get it right before I comment.

Wes “Give Claude a Chance” Roth