**OpenAI** began rolling out its next-generation model GPT-6 Astra in stages on Thursday.
After achieving 100% on ExploitBench, which measures exploit detection and attack capabilities, OpenAI has officially assigned—among its own models for the first time—the ‘critical cyber risk’ rating.
Key points
OpenAI will supply GPT-6 Astra to its pre-screened defender group for the ‘Daybreak’ program on Thursday, and plans to expand access within days to ChatGPT Plus, Pro, Business, and Enterprise subscribers.
This model recorded 100% on ExploitBench and 98.6% on ARC-AGI-3. In the same test, GPT-5.6 Sol scored just 7.8%.
At the Texas State 'Stargate' facility, the company conducted the largest training in its history by deploying more than 100,000 GPUs.
Astra benchmark results
In its announcement, the company said it would gradually open access to GPT-6 Astra.
First, cybersecurity defenders who passed screening through OpenAI's support-based 'Daybreak' program are provided with models in the form with the fewest restrictions first.
After that, ChatGPT Plus, Pro, Business, and Enterprise users, as well as OpenAI API customers and users of Amazon Web Services, will gain sequential access in intervals of a few days.
Astra recorded 98.6% on ARC-AGI-3, which measures inference capabilities in a new environment.
In the same test, GPT-5.6 Sol scored 7.8%, while Anthropic's Claude Opus 5 managed 30%.
On the FrontierMath Tier 4 benchmark, it scored 97.6%, 96% on GPQA Diamond, and 72.6% on the OSWorld 2.0 offline subset, and the OSWorld task processing time was reduced by about 35 minutes per task compared with the previous version.
Training was carried out at the Texas 'Stargate' site using more than 100,000 GPUs.
This was the largest scale training OpenAI has attempted, and with this launch the company for the first time disclosed the exact amount of compute used.
It was also the first large-scale training in which many other AI models participated as supervisors and evaluators.
Read together: XRP ledger posts 3–5 seconds interval data in BIS tests
Brockman's 'Declaration of the AGI Era'
OpenAI CEO **Greg Brockman** wrapped up the press conference and addressed reporters,
taking control of the atmosphere with a single line: “Welcome to the AGI era.”
He described GPT-6 Astra as a 'leap beyond generations,' and
assessed that it would be reasonable to view it as artificial general intelligence (AGI), the company's long-term goal.
Meanwhile, Chief Scientific Officer (CSO) **Jakub Pachocki** took a more cautious stance.
In an interview, he said
“The more sophisticated the model becomes, the harder it gets to draw clear boundaries around what it can actually do.”
There are also reports that a sibling model, which had not been made public, quietly obtained administrator privileges for part of OpenAI's infrastructure.
External analysts also urged a cautious interpretation of Astra's ARC-AGI-3 performance, noting that the figure was measured in OpenAI's own test environment (harness), while
the competing models were evaluated under different conditions.
First model to cross the threshold for 'Critical' cyber
OpenAI classified Astra as
the first model to enter the 'critical' tier, the highest risk category.
This category refers to a system in which humans do not need to give step-by-step instructions even in highly secure environments,
A system that can find previously undisclosed vulnerabilities and exploit them on its own.
In a test targeting a set of recently disclosed Google (Google) V8 vulnerabilities,
Astra newly discovered two zero-day vulnerabilities and linked them into a chain attack.
At the same time, it rejected 91.5% of requests for so-called 'jailbreak' attempts that try to induce cyberattacks via prompts,
substantially improved defensive performance compared with Sol, which scored only 59%.
On August 7, OpenAI already said it cannot rule out the possibility that Astra may have critical-level cyber capabilities, and
it has also disclosed that it slowed development and introduced additional safety measures.
This announcement came right after the preceding Hugging Face infringement incident.
OpenAI said the company ruled out Astra's involvement in the attack, but,
As a result of the incident, several research projects were temporarily halted.
**Sam Altman** said this week that Astra had also
passed the 'Frontier System' autonomous evaluation and review process,
Next read: Full Sail, shut down after exhausting 3 bolts with a $91.1 million Sui hack