OpenAI has launched a new artificial intelligence model called GPT-6 Astra. The company president, Greg Brockman, says the system represents a generational leap in capability. He also told reporters on Thursday that he personally thinks the model marks the arrival of artificial general intelligence. “It’s not unreasonable to feel that we are now in the AGI era, and I think that if you want to say this [model is] the first one, I think it’s reasonable,” Brockman said.
OpenAI defines artificial general intelligence as highly autonomous systems that outperform humans at most economically valuable work. This means the software can theoretically complete a regular job as well as, or better than, a human employee. The technology industry does not currently have a shared agreement on what exactly qualifies as artificial general intelligence, and definitions vary widely between companies. Brockman said the term no longer has a fixed trigger even inside OpenAI. “There’s no contractual AGI triggering anymore, so that’s actually not a relevant concept,” he said. He was talking about a clause in OpenAI’s old Microsoft agreement that would have changed the relationship between the two companies once AGI was declared.
The release comes three months after we reported that Anthropic reached a $965 billion valuation, making it more valuable than OpenAI at the time. Anthropic builds the Claude series of models and released Fable 5.1 and Mythos 5.1 on Tuesday, two days before Astra. The competition for enterprise customers between the two companies relies heavily on benchmark test scores.
OpenAI now claims Astra beats Anthropic’s current systems in professional work tests. The company says Astra scored 59.3% on Agents’ Last Exam. This benchmark test measures the ability of a system to complete long, complex professional tasks without human help. It covers more than 1,000 tasks across 55 professional fields and was published in June. For context, Anthropic’s Fable 5 scored 48.7% on the same test and Claude Opus 5 scored 55.2%, according to the public leaderboard. Astra is yet to appear on that leaderboard.
Astra also scored 98.6% on ARC-AGI-3, a specialised reasoning test built specifically to prevent companies from training their models on the answers beforehand. That number comes with a condition attached. OpenAI ran Astra through its own Responses API with two features switched on, Retained Reasoning and Compaction, rather than through the official test setup. Those same two settings took GPT-5.6 Sol from 7.8% to 38.3% on the same test earlier this year. On the ARC Prize Foundation’s own board, Claude Opus 5 still leads at about 30.2%. François Chollet, who co-founded the ARC Prize, has said general API settings available to every developer are fair to use as long as the settings and the running cost are reported, while setups built for one benchmark are not. So the score measures Astra plus the way OpenAI wired it up, not the model on its own.
Aidan Clark, vice president of research at OpenAI, said the company trained Astra using more than 100,000 GPUs. “It’s the first time we’ve pre-trained on more than 100,000 GPUs at our Stargate site in Texas.” Stargate is the data centre project OpenAI runs with Oracle and SoftBank. Clark also noted that older OpenAI models supervised the training of Astra.
The model is not available to the general public yet. OpenAI is giving first access to cybersecurity professionals in a restricted programme called Daybreak. The reason is not that OpenAI wants these users to hunt for faults in the model. Astra is the first system the company has classified at the critical cyber capability level under its Preparedness Framework, which means it can find unknown software flaws and build working attacks against well defended systems without step by step human instruction.
In testing, Astra found two previously unknown vulnerabilities, broke out of a browser sandbox to run commands on the machine underneath, and chained several flaws in a hardened operating system together to gain root access. OpenAI says Astra turns down 91.5% of cyber-related jailbreak attempts, against 59% for GPT-5.6 Sol.
Subscribers on Plus, Pro, Business, and Enterprise plans will get access over the coming days. The OpenAI API will also open to developers in the same timeframe, along with AWS, allowing third-party companies to build new applications using Astra. Those general versions refuse advanced cybersecurity work.
The immediate step is the enterprise security rollout. OpenAI has said it will publish Astra’s full system card at launch. That document should carry the safety and security testing behind the critical rating.
Join the discussion