Greg Brockman has spent a decade dodging the word AGI more carefully than almost anyone at OpenAI. On Wednesday he dropped the guard, saying at the GPT-6 Astra briefing that it’s “not unreasonable” to feel we’re now in the AGI era. The model went live the same morning to a small circle of Daybreak partners. Three days later, most paying ChatGPT users still can’t touch it.
I’ve spent those three days inside the system card, the developer release notes, and early Astra builds people have been running. My takeaway is blunt. The AGI framing is doing commercial work, and the parts of this launch that matter most are the ones OpenAI mentioned in passing or skipped entirely.
What actually shipped
The published specs are impressive and expensive in equal measure. Astra carries a 1,050,000-token context window split as 922k input and 128k output, a knowledge cutoff of April 30, 2026, and API pricing of $10 per million input tokens and $50 per million output. That’s a steep reset from Sol-era rates, and it lands within days of Anthropic’s Fable 5.1 at nearly identical pricing. Two labs repriced frontier intelligence upward in the same fortnight, which tells you neither fears being undercut.
The compute behind it is real. Astra is the largest pretraining run in OpenAI’s history, more than 100,000 GPUs at the Stargate site in Texas. From the early builds I’ve followed, the gains concentrate where people hoped: agentic computer use, long multi-step coding sessions, and workflow automation that would have collapsed on Sol. It’s the strongest general model anyone has shipped, by a visible margin.
Rollout is where the friction lives. Enterprise workspaces get Astra switched off by default until an admin enables it, standard Plus accounts are still phasing in as of today, and migration is nontrivial. Astra requires the Responses API for tool calling, drops the none reasoning effort setting, and rejects custom temperature, top_p, and logprobs in several configurations. Teams I’ve watched migrate spent more time reworking plumbing than testing capability.
The August pause is the story OpenAI underplayed
Here’s what the keynote didn’t lead with. In August, OpenAI halted Astra’s reinforcement learning runs after its own Preparedness evaluations flagged potential Critical cyber capabilities, then spent weeks on additional safety testing before confirming Astra as the first model rated Critical for cybersecurity under its Preparedness Framework. Parts of the framework got rewritten along the way. We covered that capability scare while it was happening, and it reads very differently now. Internal evaluations surprised the people building the model. That’s not a marketing beat, that’s a capability crossing a line nobody fully planned for.
One detail got almost no airtime. In early August, an internal Astra build produced machine-checkable proofs for ten open problems in mathematics and theoretical computer science at roughly $2,000 of inference cost. OpenAI barely mentions it in launch materials. If you want the strongest evidence for Brockman’s claim, that’s it, and it sits quietly outside the September benchmarks.
The model most customers receive is not the model that earned the Critical rating. The public version restricts advanced cyber tasks outright, refusals on dual-use prompts run dramatically higher than any previous OpenAI release, and safety monitoring can pause a conversation mid-session. Zero Data Retention is available for eligible API customers, which softens the surveillance concern a little, but the direction is unmistakable. OpenAI built something it won’t fully sell, and the gap between those two things is where the next two years of policy fights will live.
One number from the announcement got half a sentence. OpenAI is putting $1 billion in Daybreak credits behind critical infrastructure operators and cyber defenders. When a lab quietly spends that much making sure defenders hold the Critical model first, you learn exactly where it thinks the risk sits.
So is this AGI?
By OpenAI’s own definition, AGI means systems doing economically valuable work at or beyond human level. The August math results arguably satisfy that in a narrow domain, and agentic computer use is close in others. But the declaration didn’t arrive in a vacuum. Anthropic filed for its IPO first, OpenAI’s own IPO timeline stretches toward a trillion-dollar valuation, and “the company that declared AGI” reads very differently in a prospectus than “the company still shipping models.”
Where I land is simple. A model most customers can’t access, and that OpenAI itself won’t fully expose, shouldn’t carry the heaviest claim in this field’s history. The AGI-era declaration is a rhetorical act, and rhetoric this consequential deserves more scrutiny than a press briefing offers. Still, I’d bet the August pause, not the September keynote, is the moment historians will circle. Watch two things next quarter. If refusals loosen and pricing holds, OpenAI believes its own declaration. If both loosen to chase revenue, the AGI era was mostly a press release.
