SI USA Superintelligence in the USA Subscribe
Menu

Timeline Compute SIU-LDG-2026-017 SI Index +1.00

OpenAI begins rolling out GPT-6 Astra, its first model at "Critical" cyber risk

OpenAI's new flagship launches in phases, cybersecurity partners first. Epoch AI estimates it was trained on about 9.4 × 10²⁶ FLOP, the largest documented training run.

Chart of estimated training compute for notable models from 2023 to 2026, with GPT-6 Astra highest at 9.4 times 10 to the 26 FLOP, just under the 10 to the 27 line
Estimated training compute of notable models released through September 3, 2026. Chart: SI USA, from Epoch AI data / Data CC BY 4.0 · License

OpenAI said Astra is the first model to reach the “Critical” level for cybersecurity under its own Preparedness Framework, and that safeguards added after the Hugging Face incident “sufficiently minimize the risk of severe harm for release”. Companies in its Daybreak cybersecurity program got access first, followed by paid ChatGPT plans, the API and AWS.

Index treatment. Training run of 10²⁷ FLOP moves to 0.75. Capability claims at launch are OpenAI’s own and do not move the Index by themselves.