General Release

That's the actual story of GPT-5, and it isn't the benchmark. The benchmark is the press release. The story is in the architecture diagram, in one component that got about a sentence of coverage and deserves the whole article: the router.

OpenAI shipped GPT-5 as three things wearing one name. There's a fast model for ordinary requests. There's a reasoning model that thinks longer for hard ones. And there's a router — a classifier sitting in front of both, deciding in real time which one you get. You type a question. Something you cannot see and did not consent to makes a determination about how much compute your question deserves. Then you receive an answer branded GPT-5.

Hold on to that word. Deserves. We'll come back to it, and it turns out to be the whole story.

The specs, for the record: a 272,000-token context window, 88.4% on GPQA, the graduate-level science benchmark, and — as far as anyone can tell from the outside — the first frontier model placed on the European market after the AI Act's general-purpose obligations took effect on August 2nd, five days earlier. All of that is true and most of it is even impressive. None of it is the part that changes anything.

i · the router is the product

Every architecture decision is an economics decision wearing a UX costume, and this one barely bothers with the costume.

Reasoning models are expensive. They burn tokens thinking before they speak, and most of what people ask a chatbot does not require thinking — it requires retrieval, formatting, and confidence. If you hand every user unlimited access to the expensive model, you light money on fire answering questions about pasta. So you build a classifier that sends the pasta question to the cheap model and the differential equation to the expensive one, and you call the composite "GPT-5," and your cost per query drops by whatever margin the router can find.

This is a good engineering decision. I want to be clear about that, because lazy cynicism that dismisses everything equally is just a different flavor of not paying attention. Routing is the correct answer to the cost problem. Anyone building at that scale would arrive at it.

But look at what it does to the object.

Before the router, "GPT-4 said X" named something. Not something perfectly stable — there were checkpoints, there were silent updates, there was drift — but something with a referent. You could, in principle, point at the artifact. Now "GPT-5 said X" names a routing policy's outcome on one particular day for one particular phrasing. Ask the same question tomorrow with a comma moved and the router may send you somewhere else entirely. The response you got and the response your colleague got may have come from different models, and neither of you will be told.

Be precise about who that costs, because the sloppy version of this complaint is easy to dismiss and the precise version is worse. Reproducibility did not die. The API still exposes pinned model identifiers and explicit reasoning settings; a researcher who pays for access can still state exactly what they called and call it again. The evaluation literature is not homeless. What happened is narrower and uglier: the surface where hundreds of millions of people actually meet this technology stopped having a referent, while the surface where a few thousand professionals evaluate it kept one. The ability to name what you talked to became a paid feature. That is not a smaller problem than the one I was about to overstate. It's the same problem as the rest of this article, arriving early.

There's a second-order problem, too, which is that the router is a new and undisclosed failure surface. A classifier making cheap/expensive decisions under load is a classifier that can be wrong in one direction. When capacity is tight, when costs spike, when the quarter is closing — does the routing threshold move? I don't know. Nobody outside the company does, and I'm not going to dress a guess up as an accusation; an article about an unfalsifiable allocation policy has no business running on unfalsifiable claims about it. The point is the shape of the ignorance, not a charge. "The model got worse this week" is now unfalsifiable in both directions: the user can't demonstrate it, and the company can't refute it.

The argument that produces is already familiar in outline. Users report degradation. The company points out, correctly, that the underlying weights were not updated. Both parties are telling the truth. The disagreement is irresolvable by construction.

ii · what 88.4% actually measures

GPQA is a real benchmark and a good one. Rein and colleagues built it specifically to be Google-proof — questions that PhD-level experts in the domain get right about two-thirds of the time and that skilled non-experts with unrestricted web access still fail. It was designed as a hard ceiling. It was designed, explicitly, for a world where we would need to evaluate models that outperform their evaluators.

88.4% means the ceiling is gone. That's genuinely notable. It also means the benchmark has stopped being an instrument and started being a trophy.

The thing about saturated benchmarks that never makes the keynote: a score tells you about performance on the distribution the benchmark sampled. GPQA samples curated, well-posed, single-answer science questions written by people trying to be hard-but-fair. Actual use is none of those things. Actual use is ambiguous, multi-turn, badly specified, contaminated with the user's own wrong assumptions, and does not come with an answer key. The correlation between "scores 88.4% on curated hard questions" and "is reliable in the hands of someone who can't check it" is not zero, but it is nothing like one, and the marketing treats it as identity.

And now compose that with the router. Which model produced the 88.4%? I don't know, and neither do you, and that is the finding rather than a gap in my reporting. A benchmark question looks like a hard question, so the reasoning path is the obvious guess — but a guess is all it is, and the number was published without the routing disclosure that would settle it. Your query about a drug interaction, typed casually at eleven at night, may not reach the same configuration. The number on the slide may describe a place you never actually arrive, and there is no way to check.

This is not fraud. It's the ordinary drift between the eval environment and the deployment environment, and it happens every single time, and every single time the number outlives the caveat.

iii · first under the act

The European Union's general-purpose AI obligations became applicable on August 2nd, 2025. GPT-5 arrived on the 7th. Five days.

That gap is the entire governance story of this era compressed into a number you can hold in your head. The Act was drafted starting in 2021, negotiated through 2023, passed in 2024, and phased in through 2025 — call it four years of deliberate, contested, multi-stakeholder lawmaking. The thing it now governs was trained and shipped inside a fraction of that window, and the next thing will ship faster.

The obligations themselves are not nothing: technical documentation, training-data summaries, copyright policy, systemic-risk assessment for the largest models. Real requirements with real paperwork. But note the structure — they are disclosure obligations, not gating obligations. Nothing in the Act required anyone to wait. The model launches; the documentation follows; the Commission's own enforcement powers over general-purpose models don't arrive until August 2026. A full year of the most capable system yet deployed operating in the space between rules existing and rules biting.

And here is why this section belongs in an article about a router: a disclosure regime is structurally incapable of regulating a policy rather than an artifact. Technical documentation describes a model — its training data, its evaluations, its risks, its weights. The router is not a model. It's a decision procedure that can differ between your query and mine, that can be retuned on a Tuesday afternoon with nothing retrained and nothing to re-document, and that leaves no trace in any artifact the Act knows how to ask for. Europe spent four years building an instrument that makes objects legible, and the industry shipped the thing that isn't an object. You can document a model. There is no obligation anywhere in the text to document an allocation policy, and no mechanism that would notice if it changed.

I don't think this makes the Act a failure. Disclosure regimes work slowly and then matter enormously, usually in litigation nobody anticipated. But anyone reading "first major model under the EU AI Act" as reassurance has the causality backwards. It doesn't mean the model was vetted. It means the paperwork now exists.

Governance operates on the timescale of institutions. Deployment operates on the timescale of a CI run. Those two clocks are not going to converge — and worse, every framework written from here forward risks describing not just the previous generation but the previous kind of thing.

iv · who controls the on-ramp

Here's what I keep returning to.

The thing in the box is pooled human cognition. Every model of this class is built from an enormous fraction of what people have written down — arguments, corrections, jokes, textbooks, forum posts from 2009 by someone who solved your exact problem and never got thanked. It is, in the most literal sense, made of everyone. When you talk to it you are not meeting an alien intelligence; you are meeting a compression of us.

And now it is free. Genuinely free — no subscription required, hundreds of millions of people, the largest single expansion of access to that reservoir yet. That is not a small thing. The commons, opened.

But opened through a gate somebody owns.

I want to be careful with the word pulling at me here. Enclosure means something was taken — a commons people actually held, fenced off, and the holders dispossessed. That is not what happened. The free tier is strictly more access than existed the day before, and metering a free service is what every free service has always done. If this argument needs the resonance of stolen fields to supply its outrage, it isn't a good argument.

The specific wrong is narrower and entirely real, and it isn't the metering. It's the undisclosed variability. A free user and a paying user type the same question and receive different amounts of the commons, allocated by a private policy, in a quantity neither of them is told. You cannot reason about a resource whose quantity is unstated. You can't price it, budget against it, compare offerings, or contest a shortfall. You can't even know you were shortchanged — because "GPT-5" is what both of you were told you got, and both of you were told the truth. That's not a wall. Walls are visible and provoke resistance. It's a valve, and nobody organizes against a valve they can't see.

Which brings back the word from the third paragraph. Deserves. I've spent most of this article treating the router as an epistemic device, and that's the smaller half of it. It's a means test. A private classifier deciding, per person, per query, in real time, how much thinking a human being gets — with ability to pay as one of the inputs. That is cognitive means-testing, shipped at planetary scale, five days after the rules arrived, described in the launch materials as a routing improvement.

Everything else in this piece is downstream of that. The epistemic loss is downstream: you can't name what you talked to because what you talked to was an allocation, not an object. The governance gap is downstream: disclosure law reaches artifacts, and an allocation isn't one. And the distribution runs the way distributions run — the people who can afford to name their model are the same people who get the expensive one, ability to pay correlates with nearly everything else, and so the compute flows toward the advantage that already exists.

Who controls the on-ramp determines who arrives. Not who gets in — everybody gets in, that's the headline, that's the genuinely good news. But arriving where, with how much, under whose allocation, is a question with an owner, and the owner is not us.

So: a real capability jump, a real expansion of access, a real regulatory milestone, and a quiet architectural change that decides how much thought each person is issued — all in one launch, all in one week, five days after the rules arrived.

The thing everyone will argue about is the benchmark. The thing that reorganizes who gets to think with these machines shipped in a sentence, and there is currently no one with the standing to make anybody explain it.