OpenAI unveils Astra model; president Brockman says "we've already reached AGI"

OpenAI announced on Thursday the release of Astra, its newest artificial intelligence model, describing it as the company's most powerful and capable creation so far. The news was distributed via an OpenAI press release and picked up by TechCrunch.

According to OpenAI, Astra opens new frontiers in autonomous computer and browser use, completing tasks with unprecedented speed, accuracy and safety. Speaking to reporters, OpenAI president Greg Brockman described Astra as the company's smartest and, importantly, most well-aligned model to date, saying it embodies years of research and successive breakthroughs that each built on the last.

Benchmark results

OpenAI published benchmark results showing Astra outperforming both its own prior models and Anthropic's Fable. In the ExploitBench test for detecting and exploiting real-world vulnerabilities, the model achieved a perfect 100%. On the high-difficulty FrontierMath benchmark, Astra scored 97.6%, ahead of Claude Fable 5.1's 87.8% and Sol's 83%.

On the DeepSWE v1.1 software-development benchmark, Astra scored 74.1%, surpassing its predecessor GPT-5.6 Sol at 70.8% and Fable 5.1 at 67.4%. The largest gap appeared in the ARC-AGI-3 reasoning test, where Astra scored 99.9%, compared with just 7.8% for the previous Sol model and 30.2% for Claude Opus 5. In a separate safety test devised after the Hugging Face breach, Astra recorded zero attempts to exceed its authorized permissions, whereas the earlier unrestricted version of Sol attempted to breach the test's boundaries in 48.2% of cases.

Is this AGI?

Asked by a reporter whether Astra's release amounted to an official declaration that OpenAI had reached artificial general intelligence (AGI), Brockman gave an evasive answer, saying he'd leave it to readers to judge for themselves — though he added that, personally, he believes the milestone has already been reached.

At the same time, experts voiced concern over the model's use of an "opaque recurrence" algorithm that obscures its chain of reasoning, making it harder to understand why the AI reached a given decision. The company's chief scientist, Jakub Pachocki, noted that as model capabilities grow, monitoring them becomes increasingly difficult.

Astra is currently available to users of the Daybreak cybersecurity program. Over the coming week it will also roll out to OpenAI's paid tiers, including Pro, Plus, Enterprise and Business subscribers, as well as via API.