Three things change when your AI is rated ‘critical’ — and ‘AGI’ isn’t one of them

OpenAI's first 'critical' safety rating triggers new controls. The 'AGI era' claim? That's just the sticker on the box.

OpenAI just shipped a model it rates critical under its own Preparedness Framework — the top of the scale, the level that on paper is supposed to make the company nervous. Reporting around the launch calls it OpenAI’s first model to reach that tier. And then, in the same breath, the company declared the “AGI era.” Read those two facts together and you get the real story, which is not the marketing.

Start with the label, because “critical” is not a vibe. The Preparedness Framework works like a building’s fire rating. A cabin and a hospital are both “buildings,” but one legally has to have sprinklers, marked exits, and a fire marshal who can shut it down. OpenAI’s framework does the same across tracked categories — cybersecurity, model autonomy, that family of risks — on a low / medium / high / critical dial. The written rule is blunt: a model only ships if its post-mitigation score lands at medium or below, and development itself is only meant to continue at high or below. Critical is the reactor-going-critical end of that dial — the tier the policy says can gate not just release but the build.

So three concrete things change the moment a model wears that badge — and none of them is “we reached AGI.”

One: the safety case stops being optional prose. A critical rating commits OpenAI to red-team results, deployment restrictions, and access controls a medium model never triggers. What the framework doesn’t spell out is the plumbing — approval queues, usage caps, KYC-style gates, tiered access. It describes the flinch, not the specific hoop you’ll jump through. So the real integration question isn’t the benchmark. It’s whether raw capability sits behind a gate at all, because “critical for cyber” and “public API with a credit card” are hard to hold in the same hand.

Two: capability gets measured where it used to be asserted. This is the genuinely useful part — with a caveat the launch would rather you skip. arcprize.org has published Astra’s results on ARC-AGI-3, the interactive, play-a-game-you’ve-never-seen benchmark built specifically to resist memorised answers, and the improvement over prior models is the headline they’re leaning on. On the math, coding and cybersecurity suites, though, OpenAI is claiming the top spot more than it’s showing the number: several of these are asserted as “best,” not published as a score you could check line by line. So take “tops the leaderboard” as plausible and partly evidenced, and treat any clean head-to-head against Claude, Gemini or Grok as absent until someone runs them on the same test. I’m not going to invent those numbers, and neither should the launch post.

Three: the boring use case quietly becomes the strongest evidence. The Legora demo is the tell. Astra reviewed 41 financial documents and surfaced four planted errors, cutting review time by roughly 40%. Read the case study carefully before you repeat the perfect-score version: it reports Astra catching the four seeded errors, not a certified “four planted, four found, zero missed” across the whole stack. Even read conservatively, that’s the shape of work that actually pays — the kind of task where being confidently wrong costs someone their audit. If it holds outside a curated demo, it’s the paragraph practitioners should screenshot.

Which brings us to the banner. “AGI era” is doing a lot of work for a phrase OpenAI has never pinned to a threshold you could fail. Nothing in the framework says cross this ARC-AGI-3 score and it’s AGI. There’s a critical safety rating, which is a claim about danger, and a marketing claim about generality, which is a claim about vibes — and the launch quietly lets you conflate them. Don’t. A model can be genuinely more capable, genuinely worth gating, and still not be a general intelligence. All three are true here.

If you’re wiring this into anything real, treat the model name as the easy part and the access model as the actual integration. OpenAI hasn’t confirmed the exact API identifier string in the launch materials, so don’t hardcode a guess:

1
2
3
4
5
6
7
8
9
10
11
12

from openai import OpenAI
client = OpenAI()
​
# The gating, not the prompt, is what will bite you in six months.
# Confirm exact model string and access tier in docs before shipping.
response = client.chat.completions.create(
    model="gpt-6-astra",  # Placeholder; verify string and check if capability-gated
    messages=[
        {"role": "user", "content": "Reconcile these statements and flag inconsistencies."}
    ]
)
print(response.choices[0].message.content)

The honest summary: the capability leap looks real if partly unshown, the “critical” rating is the first time OpenAI’s own paperwork admits a model of its making makes it flinch, and “AGI era” is the sticker on the box. Two of those three matter. Guess which one the press release led with.