Last week OpenAI released GPT-6 Astra. Company president Greg Brockman called it a generational leap in capability. Asked whether this was the model that crossed into artificial general intelligence, he said he thought it might be about this model, then closed the briefing by welcoming the room to the AGI era.
Declaring AGI is the oldest play in the book and I have been sceptical about it for years. Something else at that same briefing got much less attention, and it is the reason I think the claim holds this time.
Astra is the first OpenAI model rated critical for cybersecurity under the company's own Preparedness Framework. It scored 100% on ExploitBench and found two zero-day vulnerabilities in Google's V8 engine, the JavaScript engine behind Chrome.
Brockman confirmed it is the first system OpenAI has rated capable of hacking well-protected systems without human guidance. The most capable version is being held back to a small group of trusted testers.
Marc Andreessen posted on X in April that AGI is already here and just not evenly distributed yet. He took the line from William Gibson. I took it from him, because it describes this moment better than any benchmark does.
The definition problem has not gone away
OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. Brockman told Fortune it has arrived in bits and pieces rather than in one moment, which matches how these shifts tend to feel from inside them.
Take that definition seriously, allow for the physical limits of something that lives in a computer, and we are past it. Astra is trained to use a computer the way a person does, moving across browsers, spreadsheets and desktop applications to finish multi-step work.
OpenAI says it handles financial modelling, tax preparation and video game development with minimal oversight. On ARC-AGI-3, the benchmark built to test reasoning on problems a model has never seen, it scored above 99.9%.
Benchmarks have been gamed before, and ARC has been revised twice because models kept saturating it. Sam Altman stepped back from the term in August 2025, saying it means different things to different people. The trajectory still runs one way.
Yann LeCun continues to argue there is no such thing as general intelligence. Demis Hassabis, who called that incorrect, wrote in July that we are in the foothills of the singularity and AGI is a few years out. The one AI leader with a Nobel Prize thinks Brockman is early.
Partner message: Get 50% off your first month of Hedra
Most agents stop at a document. Hedra Agent 2 keeps going and makes the thing.
It works in a Space, a persistent workspace the agent reads from and writes to as it goes. Give it a starting point and it asks clarifying questions, proposes a plan, researches the web, and writes the brief or script onto the canvas in front of you.
It pulls from Notion, writes into a Google Doc, runs code in a sandbox, saves repeat requests as Skills, and runs whole workflows on a schedule. Then it picks the right media model and produces the finished video, image or audio from the same session.
Try it at hedra.com/agentic using code agentic50 for 50% off your first month.
AGI arrived as a service tiers
Astra went first to enterprise customers already inside OpenAI's Daybreak Access programme. Plus, Pro, Business and Enterprise subscribers get it in the coming days with cyber guardrails added. The full model stays with a small group of vetted testers.
Astra was OpenAI's largest training run, the first on more than 100,000 GPUs at Stargate in Texas. It is also the first OpenAI release where earlier models supervised the training. Models training models is the mechanism behind every intelligence explosion scenario.
Aidan Clark at OpenAI mentioned it as a technical footnote. Nick Bostrom imagined a single system crossing a single threshold while the world reorganised around it. What exists instead is a capability rationed by who pays, who is trusted and who government has reviewed.
Government has reviewed it. Brockman said OpenAI ran its testing process with the US administration and nobody came back asking for changes. The June executive order gives agencies up to 30 days of early access and states that it authorises no licensing or pre-clearance.
Andrej Karpathy calls the result jagged intelligence. A model that finds zero-days in Chrome can still fail a task a competent teenager would manage, and you cannot tell in advance which side of the line you are standing on.
The cost is landing on the youngest workers
The Federal Reserve Bank of New York put unemployment among recent college graduates at 5.6% in the fourth quarter of 2025, above the national rate. Underemployment, the share of graduates in jobs that do not need a degree, reached 42.5%.
Stanford's Digital Economy Lab found a 13% relative decline in employment for US workers aged 22 to 25 in the occupations most exposed to generative AI. Older workers in the same occupations saw no such fall.
Senator Mark Warner has warned that graduate unemployment could reach 30 to 35% before 2028. ServiceNow chief executive Bill McDermott gave a near identical forecast on CNBC a few weeks earlier.
The counter-evidence deserves a hearing. ZipRecruiter found the share of graduates landing work within three months rose from 63% to 77% year on year, and McKinsey plans to lift junior hiring in North America by 12%. Hiring fell broadly after 2022.
If the first rung goes, the pipeline goes with it. The senior analyst of 2032 is the junior analyst nobody hired in 2026. A cohort that enters work underemployed and in debt becomes a drag on consumption that the rest of the economy will eventually feel.
