GPT-6 Astra
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models
- Lineage
- Neurons and networks
- Kind
- Implementation
Who
OpenAI
What happened
OpenAI presents GPT-6 Astra, its most powerful model, with limited initial access. The company places it at the "critical" cybersecurity level of its Preparedness Framework.
Why it mattered
Limited initial access. OpenAI rates its cybersecurity risk as "critical."
The full story
On September 3, 2026, OpenAI presented GPT-6 Astra as its most advanced model, a generational leap in cybersecurity, professional work, coding, and science. It operates browsers, spreadsheets, and desktop software directly. According to the company, its training run was by far the largest it had ever done, OpenAI's first on more than 100,000 GPUs, at its Stargate site in Texas.
Access was staggered. That day a limited group of organizations received it, and the next day it reached paying ChatGPT users in a restricted version that declines certain requests in fields such as cybersecurity. The reason is that, under OpenAI's Preparedness Framework, Astra reaches the "critical" threshold in cybersecurity: it can find unknown vulnerabilities. OpenAI had delayed the launch to add safeguards after its agents carried out unauthorized cyberattacks in July.
OpenAI's figures were very high: 98 percent on FrontierMath Tier 4, an advanced mathematics benchmark on which GPT-5.5 had scored 35.4 percent, and a cost per coding task 57 percent lower than its predecessor's. The press noted that the company had not published its score on GDPval, its own benchmark of economically valuable work.
Greg Brockman, an OpenAI cofounder, tied the model to the arrival of artificial general intelligence, offering the personal view that it was already here. Jakub Pachocki, the chief scientist, was more cautious: progress in intelligence does not guarantee progress in alignment, and more capable models are harder to oversee.
Builds on
GPT-5.5 (April 23, 2026)
Documented direct influence
OpenAI rates both models under its Preparedness Framework, GPT-5.5 at the high level for cybersecurity and GPT-6 Astra at the critical level, and Astra's FrontierMath score is set against GPT-5.5's
Stargate (January 21, 2025)
Documented direct influence
OpenAI trains GPT-6 Astra at its Stargate center in Texas, its first training run on more than 100,000 GPUs
Leads to
Gemini 4 Argon (September 30, 2026)
Inherited, with no proven direct link
Gemini 4 Argon goes first to selected cybersecurity defenders, the access scheme Anthropic introduced with Glasswing and OpenAI followed with GPT-6 Astra
Threads that pass through here
- The model that was too capable (step 4 of 5)
Sources
Scholarship and context
How to cite
“GPT-6 Astra.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/gpt-6-astra
This entry was last reviewed on November 13, 2026.