OpenAI Drops GPT-5.5 Seven Weeks After 5.4. The Previous Four Were Just Practice, Apparently.

I want you to do something for me. Open a new tab. Google “GPT-5.4 review.” Read the first paragraph. Then close that tab and come back here, because according to OpenAI, it’s already over. GPT-5.4 — the model they released in early March, the one they called their most capable yet, the one enterprise contracts were probably signed over — has been quietly lapped by its own successor, seven weeks later.

Welcome to the GPT-5 era, where every six to eight weeks is apparently a geological epoch.

On Thursday, OpenAI announced GPT-5.5, which it is calling its “smartest and most intuitive to use model.” Greg Brockman, OpenAI’s president, described it as “a new class of intelligence” and “a big step towards more agentic and intuitive computing.” This is a direct quote. He said this. Out loud. Into a microphone. About the fifth-point-fifth version of the fifth generation of a model, which itself came seven weeks after he presumably said something very similar about GPT-5.4.

I’ve covered a lot of OpenAI announcements. I’ve watched them sunset DALL-E for something called Maskingtape-Alpha. I’ve tracked SoftBank borrow 0 billion to hand them 0 billion. But “new class of intelligence” for GPT-5.5 is a phrase doing extraordinary work. That phrase is lifting boulders with its bare hands.

The Version Number Timeline Is Having a Crisis

Let’s take a moment to appreciate the GPT-5 release calendar, which reads less like a product roadmap and more like someone counting too fast in a dream:

• GPT-5.1: Post-training iteration. “Smarter.”
• GPT-5.2: Post-training iteration. “Even smarter.”
• GPT-5.3: Post-training iteration. “Still getting smarter.”
• GPT-5.4: Post-training iteration. “Smartest and most capable yet.” Also: first with native computer-use capabilities. Big deal, genuinely.
• GPT-5.5: First fully retrained base model since GPT-4.5. Actually a new model. “New class of intelligence.”

Here’s the detail worth sitting with: GPT-5.5 is legitimately different from its predecessors. GPT-5.1 through 5.4 were all post-training iterations on the same base model — refinements stacked on top of refinements. GPT-5.5 is an actual full retrain. The first one since GPT-4.5. Which means OpenAI spent the last several months shipping incremental updates under incrementally confident names, building to this moment where they could finally drop the real thing and call it... 5.5.

Not GPT-6. Not even GPT-5.5 Ultra. Just 5.5. After five point four. Which followed five point three. I have a lot of feelings about this. Most of them are decimal points.

The Super App Cometh (It Always Cometh)

OpenAI’s real ambition, hiding behind the version number drama, is more interesting than the number itself. GPT-5.5 is being positioned as the foundation of an AI “super app” — a unified platform combining ChatGPT, Codex, browser capabilities, and agentic task completion into one thing you never close.

This is Silicon Valley’s oldest idea wearing new clothes. WeChat did it. Instagram tried it. Every major platform has, at some point, looked at its own success and thought: what if we were also everything else? Now it’s OpenAI’s turn, and to their credit, they have more material to work with than most. GPT-5.5 can analyze data, debug code, operate software, research the web, and produce documents — all in one go, with less hand-holding required.

The benchmarks are real: 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro. These are not numbers OpenAI invented this morning. The model is genuinely capable at multi-step agentic work, and the super app vision — chaotic version numbering and all — isn’t incoherent.

But the super app dream requires one thing above all else: stability. Users need to trust that what they’re building around or relying on will still be there, roughly the same, next month. OpenAI has answered this very reasonable concern with a new flagship model every seven weeks. They’ve already built an app store. They’ve built an agent layer. They’re building everything — fast — and the super app will hold it all together, once the version number stops moving long enough for someone to trust it.

The Price of a “New Class of Intelligence” Is Exactly 2x the Last Class

GPT-5.5 is priced at 2x the API cost of GPT-5.4. OpenAI notes, helpfully, that it is “more token efficient,” which is the kind of sentence that means different things to different people. To OpenAI, it means: you’ll use fewer tokens to accomplish the same thing, so even though each token costs twice as much, the math works out. To enterprise customers currently mid-sprint on a GPT-5.4 integration, it means: time to update the spreadsheet.

Separately, OpenAI told developers that the model requires “different safeguards” before being deployed to the API — a tactful way of saying the model that can operate your software autonomously and research things online will not be available to anyone with a credit card immediately. The API is coming “very soon.” In OpenAI time, this could mean anything from next Tuesday to before the tomatoes on my windowsill ripen.

A Brief Genuine Moment, Then We Return to Our Regularly Scheduled Snark

I want to say this plainly: GPT-5.5 being the first fully retrained model since GPT-4.5 is a meaningful technical milestone. It’s not just another coat of paint. OpenAI built something substantively new here — the first flagship positioned primarily as an agent runtime rather than a chat model — and that distinction actually matters.

I’ve written before about whether OpenAI is actually in trouble, and the honest answer, then and now, is: probably not. There’s a world where GPT-5.5 is the inflection point historians point to when explaining how ChatGPT became the thing people actually do their jobs inside. Brockman’s “new class of intelligence” line might age fine. It might be the understatement of the year.

It’s just difficult to feel that gravity when it arrives seven weeks after the last announcement, costs twice as much as the thing you just set up, and shares a version number so close to its predecessors that you have to squint to confirm it’s different. OpenAI is building toward something genuinely ambitious — a platform that knows you, works for you, lives in everything you do. And they are building it in the most OpenAI way possible: shipping so fast that each launch quietly obsoletes the last, asking enterprise customers to build on a foundation that’s actively being replaced underneath them.

GPT-5.5. A new class of intelligence. Real work starts now.

GPT-5.6 will be out before the tomatoes are ready. I’m watching them.