OpenAI has officially launched GPT-6 Astra as its newest flagship model tailored for complex autonomous workflows. The system is specifically engineered for end-to-end coding, computer use, and advanced offensive as well as defensive cybersecurity operations. With this release, OpenAI is targeting enterprise developers looking to delegate entire multi-step software tasks to agentic systems rather than relying on standard conversational interfaces.
A notable milestone accompanying the release is the evaluation detailed in the documentation. OpenAI classified GPT-6 Astra's cybersecurity capabilities as 'Critical' for the first time in its official system card. This safety rating highlights the model's proficiency in analyzing source code, identifying software vulnerabilities, and executing automated defenses, while also triggering tighter governance requirements for deployment in production environments.
On the economic side, OpenAI has priced the API access at 10 dollars per million input tokens and 50 dollars per million output tokens. In comparison to its predecessor GPT-5.6 Sol, Astra operates with up to 70 percent greater token efficiency across agentic workloads. This optimization significantly lowers operational costs when orchestrating persistent autonomous agents that execute numerous computational steps to solve a single assignment.
Performance measurements on benchmark evaluations have also sparked significant interest across the technical community. On ARC-AGI-3, GPT-6 Astra achieved a score of 99.9 percent when evaluated using the vendor's dedicated provider adapter harness. When measured against the standard evaluation harness, however, the score dropped to 62.7 percent, highlighting the substantial impact that harness configurations have on benchmark results.
The model launch coincides with an ongoing technical debate surrounding interpretability in autonomous agents. Industry researchers point out that modern systems increasingly rely on internal latent vector exchanges during recursive operational loops rather than generating legible token chains. This shift poses new operational challenges for safety teams tasked with prompt auditing and real-time monitoring before letting autonomous software run freely.