OpenAI introduced GPT-6 Astra — a new flagship model that has made significant gains in programming, mathematics, computer use, and cybersecurity.
The main figure from the release — is 99.9% on ARC-AGI-3. Humans score 100% on this test, while GPT-5.6 Sol — only 7.8%.
But there is an important detail: Astra achieved 99.9% in a special testing mode with state preserved between requests. In the standard environment, the result was 62.7%.
What Astra Can Do
The model has learned to work much better not only with text, but also with computers:
-
controls browsers and applications;
-
fills out forms and works with documents;
-
writes and tests code;
-
builds websites;
-
performs long chains of tasks without constant human prompts;
-
helps with scientific and mathematical tasks.
In OSWorld 2.0, Astra scored 72.6% compared with 65.7% for GPT-5.6 Sol, while completing tasks around 47% faster.
Especially Powerful in Cybersecurity
Astra scored 100% on ExploitBench compared with 78.5% for GPT-5.6 Sol. The model is capable of independently finding vulnerabilities and creating working exploits.
Because of this, OpenAI is currently restricting the most powerful cyber capabilities: verified security professionals get access to them first.
And What About AGI?
OpenAI President Greg Brockman said after the presentation: «Welcome to the AGI era». The company considers Astra the first model to come close enough to its definition of artificial general intelligence.
But the results themselves show a more interesting picture: Astra is extremely strong in certain complex tasks, while its capabilities depend heavily on exactly how the test is conducted.
When Will It Be Available to Try?
Astra is currently being gradually rolled out to a limited number of organizations. In the coming days, the model is expected to become available to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the API and AWS.
For those who use AI tools for work, the most interesting part here is not even the benchmark score. Astra is getting much better at tasks where previously you had to open websites yourself, switch between programs, check the result, and give the model the next command. Now it can take over a significant part of this routine.
