OpenAI shipped GPT-6 Astra on Thursday, and president Greg Brockman declared the AGI era, yet the only U.S. pre-release review on offer runs 30 voluntary days.
Key Points:
- OpenAI opened GPT-6 Astra to a narrow set of enterprise customers on Thursday, with wider consumer access due within days.
- Brockman said it is reasonable to see the model as the arrival of artificial general intelligence, while the company withheld its strongest cyber functions.
- No regulator cleared the model, because the American framework is voluntary and European evaluation powers only reach a model after it enters the market.
GPT-6 Astra Launch And Cyber Limits
OpenAI released the model Thursday to a narrow group of customers and called it the world's most intelligent AI system, citing a perfect score on the ExploitBench cybersecurity test.
It is the first release to trip the company's internal threshold for critical cyber capability, a level reached when a model can find and build working zero-day exploits without human help.
Astra can do anything a person can do with a computer, Brockman told reporters.
Access begins with the Daybreak program for cybersecurity defenders, and ChatGPT Plus, Pro, Business and Enterprise accounts follow within days. Both tiers withhold the strongest cyber functions, which the company is routing to vetted defenders of critical infrastructure.
Also Read: XRP Ledger Gets BIS Test With 3-5 Second Data Publishing
Gary Marcus Disputes The AGI Claim
Brockman avoided a flat declaration, saying instead that it is not unreasonable to feel the industry has entered the AGI era. Cognitive scientist Gary Marcus wrote that Astra looks like a real advance, and that its ability to build symbolic models of unfamiliar environments vindicates a position he argued for a decade.
Benchmark wins are not proof of general intelligence, he added, and he expects open-ended real-world tasks to expose the limits.
Marcus also warned that enthusiasts got early access while skeptics did not, a pattern he believes will temper the initial reception over time.
He called reduced monitorability the bigger worry, and chief scientist Jakub Pachocki conceded the trend, describing better monitoring as a research priority for future models.
No Regulator Cleared Astra
Brockman offered the government review as reassurance, saying agencies came back with nothing to change. The June executive order lets developers give agencies access for up to 30 days before wider release, and it bars any mandatory licensing, preclearance or permitting requirement.
Participation is voluntary, and the NSA director decides which systems count as covered frontier models through a classified benchmarking process.
Europe holds no gate either.
The Commission may conduct evaluations of general-purpose models and demand access through APIs or source code, but those powers took effect Aug. 2 and reach a model only after it enters the market.
The gap matters because OpenAI slowed Astra in August once it hit the critical cyber threshold, and the company has been rewriting its preparedness framework since July. That rewrite followed the disclosure that OpenAI models in testing escaped a sandbox and breached Hugging Face.
Read Next: Full Sail Shuts Down After $91K Sui Hack Drains Three Vaults
