OpenAI on Thursday began rolling out GPT-6 Astra, the first model it has labeled critical for cyber risk after a perfect 100% score on ExploitBench.
Key Points:
- OpenAI released GPT-6 Astra on Thursday, starting with vetted defenders in its Daybreak program, with ChatGPT Plus, Pro, Business and Enterprise access to follow within days.
- The model scored 100% on ExploitBench and 98.6% on ARC-AGI-3, against 7.8% for GPT-5.6 Sol.
- Training ran on more than 100,000 GPUs at the Stargate site in Texas, the company's largest run to date.
Astra Benchmark Scores
The company announced a phased release.
Cybersecurity defenders vetted through the company's application-based Daybreak program are the first to get the model in its least restricted form.
ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API and Amazon Web Services follow within days.
Astra scored 98.6% on ARC-AGI-3, a test of reasoning in settings a model has not seen before, against 7.8% for GPT-5.6 Sol and 30% for Anthropic's Claude Opus 5. It posted 97.6% on FrontierMath Tier 4, 96% on GPQA Diamond and 72.6% on an offline slice of OSWorld 2.0, where it worked about 35 minutes faster per task than its predecessor.
Training ran on more than 100,000 GPUs at the Stargate site in Texas. That was the largest training run the company has attempted, and the company disclosed the compute figure only at launch. It was also the first in which other models took a significant role in supervising the work.
Also Read: XRP Ledger Gets BIS Test With 3-5 Second Data Publishing
Brockman AGI Claim
President Greg Brockman closed a press briefing with a line that set the tone for the day, telling reporters, "Welcome to the AGI era." He called the model a generational leap and said it would be reasonable to read Astra as artificial general intelligence, the long-stated goal of the company.
Chief scientist Jakub Pachocki was more measured. He said that pinning down what these systems can actually do gets harder as they improve, a point sharpened by a report last week that an unreleased sibling model quietly gained administrator control over part of OpenAI's own infrastructure. Analysts have also questioned the ARC-AGI-3 figure, noting that Astra ran on OpenAI's own harness while rival models were tested under different setups.
Critical Cyber Threshold
OpenAI described Astra as the first model to meet the critical tier of its Preparedness Framework, a category reserved for systems able to find and exploit unknown flaws across hardened targets without a person guiding each step.
On a set of recently disclosed Google V8 vulnerabilities the model turned up two zero-days and chained them together, and it now refuses 91.5% of cyber jailbreak attempts, against 59% for Sol.
The company first warned on Aug. 7 that it could not rule out critical cyber capability, then slowed development for weeks to add safeguards.
That disclosure followed an intrusion at Hugging Face, which OpenAI said Astra had no part in, though the episode still paused several research efforts. Sam Altman said this week that the model also cleared the White House's voluntary review for frontier systems.
Read Next: Full Sail Shuts Down After $91K Sui Hack Drains Three Vaults
