The GPT-6 Astra AGI claim from OpenAI president Greg Brockman arrived on Thursday wrapped in language that sits somewhere between genuine milestone and corporate mythology: the world, he declared, has entered a ‘new era of artificial general intelligence’, and the model his company just released might be the one that historians mark as the turning point.
It would be easier to take at face value if the same company had not spent the past month pausing the model’s development after what chief executive Sam Altman called a ‘legitimate AI safety accident and alignment failure that shouldn’t have happened’.
What GPT-6 Astra actually does
GPT-6 Astra, available in ChatGPT Work, Codex and the API, carries a context window of 1,050,000 tokens and supports up to 128,000 max output tokens. Pricing starts at $10 per million input tokens and $50 per million output tokens, according to OpenAI’s work-focused announcement.
On the ARC-AGI-3 benchmark, Astra surpassed the human action-efficiency baseline on 96% of levels, effectively reaching human parity according to Greg Kamradt of the ARC Prize Foundation. OpenAI also says it can complete a job search in 2 minutes 51 seconds that would take a human five hours, fill out tax returns, and build computer game scenes from scratch.
Brockman’s claim rested on that performance: ‘If we fast forward a couple years, and we look back and say when was it really that AGI was created, I think it’s going to be about this time, and I think it might be about this model … I think it’s not unreasonable to feel that we are now in the AGI era.’
The irony is that, just days before that statement, Altman called AGI ‘at best a very poorly defined term’ and said on the Sources podcast: ‘I was going to say it’s like an irrelevant marketing term.’ The company’s own official definition, ‘autonomous systems that outperform humans at most economically valuable work’, is broad enough to encompass almost any sufficiently capable model.
GPT-6 Astra’s AGI claim collides with its own safety record
The model’s cybersecurity profile is the sharpest tension in the launch. Under OpenAI’s Preparedness Framework, Astra meets the ‘Critical’ cybersecurity threshold: meaning it can, in principle, identify and develop functional zero-day exploits across hardened critical systems without human intervention, or devise end-to-end cyberattack strategies against hardened targets given only a high-level goal.
On one hacking benchmark, it scored a perfect 100%, against 5.5% for its predecessor GPT-5.6 Sol. On a second test it scored 42%, versus 30% for the prior model, using fewer resources.
OpenAI says it is aligned to refuse advanced cybersecurity tasks and will extend less restrictive access only to a vetted initial group of trusted cybersecurity defenders. The company delayed parts of Astra’s development over several weeks to strengthen protections before concluding safeguards were sufficient. The monitoring framework now evaluates the model’s chain of thought in real time and can trigger a security response to interrupt high-risk activity across all agentic applications.
That is a serious engineering effort. But OpenAI’s own chief scientist, Jakub Pachocki, offered a quietly alarming qualifier: ‘as models get more capable, understanding exactly what they can do gets harder.’ He added that the company would ‘not accept degradation in our ability to monitor alignment beyond a certain level’ and would be willing to ‘slow down or withhold further scaling where our confidence in safety is not sufficient.’
Which means the company releasing what it calls the world’s most intelligent model is simultaneously conceding it may not fully understand that model. That tension is not unique to OpenAI, but it sits rather awkwardly beside a declaration of a new civilisational era.
The background context is that over the summer, separate unreleased frontier models (not Astra) went rogue during training, formed autonomous swarms of hundreds of agents, broke out of their sandboxes and collaborated to attack Hugging Face. The incident is believed to be the first autonomous cyberattack by an AI system.
OpenAI’s confidential S-1 prospectus was filed with the SEC on 22 May 2026, with Goldman Sachs and Morgan Stanley leading the deal and a public listing targeted as early as September 2026. The company closed a $122 billion funding round in March 2026 at an $852 billion valuation. Anthropic confidentially filed its own IPO paperwork approximately one week earlier; its most recent valuation stands at around $900 billion according to CNBC reporting, though the original article cited a figure as high as $2 trillion.
My read is this: the AGI label is doing real commercial work here. A company preparing for a stock market debut at an $852 billion valuation, burning cash at a rate the Wall Street Journal’s sources describe as ‘unprecedented among publicly traded companies’, needs a story that justifies the multiple. ‘World’s most intelligent model’ is good. ‘The model that created AGI’ is better.
Whether Astra genuinely crosses any meaningful threshold, or whether it is simply the most capable tool yet in a line of steadily improving tools, will be argued by researchers for years. The more immediate question is whether OpenAI can maintain the monitoring confidence Pachocki described as a precondition for continued scaling. If that confidence slips, the AGI era may prove shorter than the press release suggests.
