AI Signal 95
OpenAI releases GPT-6 Astra with computer-use agents, million-token context, and critical cybersecurity classification
↗
OpenAI released GPT-6 Astra, a model that performs multi-step tasks through graphical interfaces, supports up to one million tokens of context, and is the company's first model classified at the critical cybersecurity capability level under its Preparedness Framework.
GPT-6 Astra moves OpenAI's models from generating responses toward directly executing multi-step tasks in software environments, including filling forms, updating records, and installing software through graphical interfaces. The model's critical cybersecurity classification and its demonstrated ability to discover previously unknown vulnerabilities in unsafeguarded testing raise monitoring concerns, as OpenAI found Astra's written reasoning harder to monitor than its predecessor.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
GPT-6 Astra can interact with graphical interfaces to perform multi-step tasks like filling forms, updating CRM records, and installing software, scoring 72.6% on OSWorld 2.0 compared to 65.7% for GPT-5.6 Sol.
The model supports up to one million tokens of context with a 96.3% MRCR score in the 512K-to-1M range, and introduces an experimental context mechanism in Codex that maintains notes across context windows with searchable previous windows.
Astra is OpenAI's first model classified at the critical cybersecurity capability level, having discovered two previously unknown vulnerabilities in testing without production safeguards, while the production version restricts advanced offensive tasks and OpenAI plans broader defensive capabilities through its Daybreak program.
THE CLUSTER