OpenAI shipped GPT-5.5 and the real story is not the benchmarks. It is the super app.

OpenAI shipped GPT-5.5 on April 23, 2026. I spent a few hours reading the announcement, the TechCrunch writeup, and the Latent.Space roundup, and I want to get my read on paper before the takes start crowding each other out.

The short version is that the model itself is a legitimate step up, but the more interesting thing is what OpenAI is signaling about where the product is going.

What OpenAI actually said

The company's own framing is that GPT-5.5 is "a new class of intelligence for real work." TechCrunch quotes OpenAI calling it their "smartest and most intuitive to use model" yet, which is the kind of sentence you are supposed to roll your eyes at. I did, briefly, and then I looked at the rest of the announcement.

Greg Brockman put it in words I actually liked: "This model is a real step forward towards the kind of computing that we expect in the future. It's a faster, sharper thinker for fewer tokens compared to something like 5.4." The "fewer tokens" part is not marketing filler. OpenAI says 5.5 matches 5.4 on per-token latency in production serving while operating at a higher level of intelligence, and uses significantly fewer tokens to complete the same Codex tasks. That is a quieter claim than a benchmark jump, and in practice it is the one that shows up on your invoice.

Mark Chen said 5.5 is better at navigating computer work and "shows meaningful gains on scientific and technical research workflows." Chen does not usually oversell, so I take that seriously.

The benchmark card

I am usually cold on benchmark announcements, but these are worth writing down because they tell you where OpenAI is pointing the model.

• Terminal-Bench 2.0: 82.7
• GDPval: 84.9
• OSWorld-Verified: 78.7
• CyberGym: 81.8
• FrontierMath Tier 1-3: 51.7

Two things stand out. First, OSWorld-Verified is a computer-use benchmark. A model scoring in the high 70s on realistic desktop tasks is a different category of thing than a model that writes pretty essays. Second, CyberGym at 81.8 is high enough that OpenAI is treating 5.5 as High under the biological, chemical, and cybersecurity categories of its Preparedness Framework. That is a capability classification with real safety scaffolding attached to it, not a press release line.

FrontierMath at 51.7 is not a ceiling result but it is the kind of number that shifts what you can hand to an AI and walk away. A year ago that same benchmark was in the single digits for frontier models.

Pricing, context, and hardware

API pricing is $5 per million input tokens and $30 per million output tokens for 5.5. GPT-5.5 Pro is $30 in, $180 out. Context window is 1M in the API and 400K in Codex. Availability is live in Codex and ChatGPT now, with API access coming soon. In ChatGPT, 5.5 goes to Plus, Pro, Business, and Enterprise; 5.5 Pro stays locked to Pro, Business, and Enterprise.

The infrastructure footnote is worth paying attention to. OpenAI says the model was co-designed for and served on NVIDIA GB200 and GB300 NVL72 systems. That is a specific hardware dependency, and it is the kind of detail that tells you the model's performance envelope is built around a particular generation of Nvidia silicon. If you are trying to think about whether OpenAI is chip-constrained heading into the next cycle, this is a relevant data point. They are not hedging away from Nvidia.

What Latent.Space noticed

The Latent.Space roundup called 5.5 "the day's dominant release" and, more usefully, pointed out that the main story is not the benchmark card. It is long-horizon execution, stronger computer-use behavior, and better token efficiency.

That matches what Brockman and Chen said, and it matches my read of the benchmarks. Headline scores on math problems get the press, but agents that can actually hold a task together for an hour and touch a browser and a spreadsheet and a terminal without falling apart are the thing that changes how work gets done.

Latent also flagged the Codex upgrades that shipped alongside the model. Browser control. Sheets and Slides. Docs and PDFs. OS-wide dictation. Auto-review mode. That is not one product feature. That is Codex turning into something much larger than a coding assistant.

The super app is the actual story

Here is the line from TechCrunch that I keep circling back to. OpenAI is framing GPT-5.5 as part of its path toward a "super app" that combines ChatGPT, Codex, and AI browser capabilities into one unified service.

That framing is the real release today. Not the benchmarks. Not the pricing. The fact that OpenAI is openly using the words "super app" about its own roadmap.

Think about what the current OpenAI product surface actually is. ChatGPT is the consumer and prosumer front door. Codex is the developer agent. The AI browser is the computer-use layer. Each one has been maturing on its own track. Today's announcement reads as OpenAI deciding those tracks converge, and that GPT-5.5 is the model with enough computer-use capability and enough token efficiency to sit underneath all three at the same time without costs or latency falling apart.

The tell that this is real, and not just a slide, is the internal number OpenAI slipped in. More than 85 percent of the company uses Codex every week. When a company of OpenAI's size runs its own coding through its own agent at that kind of saturation, it is because the product has crossed from demo into dependency. That is the version of Codex they are now trying to externalize as the base of a super app.

Latent.Space frames the strategic story the same way. The Codex upgrades are not a side release. They are the foundation. The model is what makes the foundation strong enough to carry the rest.

What I think this means for builders

I want to be careful not to over-extrapolate from one launch day. But a few things feel fairly settled after reading the material.

Token efficiency is now a first-class selling point. OpenAI is calling it out on the announcement page and Brockman is leading with it in press. The cost-per-task delta, not the cost-per-token sticker, is the number that matters for anyone deploying these models at scale. If 5.5 uses significantly fewer tokens to finish the same Codex task as 5.4, the effective price cut is larger than the surface pricing suggests.

Computer use is leaving beta. A 78.7 on OSWorld-Verified plus OS-wide dictation plus browser control inside Codex adds up to a model OpenAI expects people to hand real desktop work to. That changes the threat model for anyone running internal data through these tools, and it changes the opportunity model for anyone building automation on top of them. Both sides of that are going to need new guardrails.

The super app framing is going to pull the industry. Once OpenAI starts pitching a unified assistant that codes, browses, drafts, and researches inside one product, every other vendor has to decide whether they compete at that level, specialize beneath it, or integrate with it. I do not think any of those are bad answers. I think most teams have not decided which one they are picking.

The thing I am still watching

I am still watching the Preparedness Framework piece. OpenAI classifying 5.5 as High in biological, chemical, and cybersecurity capability is not a small footnote. It means the model crossed a threshold that triggers specific internal safeguards before release. That is the framework working as designed, but it is also the first time I remember a frontier model shipping to Plus and Business tiers with that kind of capability rating attached. The question of how those safeguards hold up in the wild, across API access and Codex workflows, is going to matter more than any benchmark score they released today.

Everything else in this announcement is upside. That part is the one worth watching with patience.