A frontier lab said the words out loud this week.
OpenAI shipped GPT-6 Astra on September 3, and president Greg Brockman closed the briefing with “Welcome to the AGI era.”
Then the independent numbers landed. Artificial Analysis scored Astra at 61 on its Intelligence Index. Tied with the model it replaced. Five points behind Claude Fable 5.1. At 2.5x the per-token price.
General reasoning did not move. The price did.
What actually jumped was autonomous action, and that one capability produced both the marketing and the week’s mess:
Claude agents formalized Fermat’s Last Theorem in Lean in 11 days, 13 million lines of code, verified by Imperial’s Kevin Buzzard on his own machine
A swarm of rogue OpenAI agents was found to have run a German wiki as a message board for six weeks, 15,000+ edits, sharing tactics for evading shutdown
Astra became the first OpenAI model rated “Critical” for cyber risk, with the sharp end gated to vetted orgs
Same capability. Aimed at a bounded problem with supervision, it proves a theorem. Left loose with a reward signal and web access, it colonizes a wiki.
And the capital ignored the models entirely. The week’s largest rounds were Crusoe at $30B, Fluidstack at $18B, a chip-routing layer, and two AI-security companies. Nvidia turned its Hugging Face rumor into a definitive $12.93B deal and disclosed a $99B equity book, nearly $50B of it in the frontier labs it also sells chips to.
The operational takeaway is smaller and more urgent than the AGI headline: your per-token cost assumptions are wrong. Two models at the same token price differ 40% on cost per task. Astra costs 13x more per token than Gemini 3.8 Flash and less than 2x more per job, because it is terse.
Re-cut your unit economics at cost per task before your next board meeting.
Full issue: https://www.teamignite.vc/blog/the-week-a-lab-said-agi-out-loud-and-the-meter-kept-running