OpenAI Releases GPT-6 Astra, Achieving SOTA in Professional Tasks and Improving Cost Efficiency

OpenAI has released GPT-6 Astra and made it available via ChatGPT Work, Codex, and the API. The model achieves SOTA (State of the Art) performance in computer use, software engineering, cybersecurity, and science. It can write code and perform tasks without data preparation or custom integration for compatibility with existing applications.

Cost Efficiency and Performance

Astra is trained to complete tasks with fewer tokens and retries, reducing cost per task. It holds most of the cost-efficiency frontier in professional task evaluations such as Terminal Bench 4.0 and the Artificial Analysis Intelligence Index. Pricing is $10 per million input tokens and $50 per million output tokens.

Enterprise Adoption and Benchmarks

Multiple companies have adopted Astra and confirmed performance improvements.

• Cognition: Achieved SOTA on internal benchmarks through integration with the Devin harness.
• Databricks: Achieved SOTA on the OfficeQA Pro benchmark and improved cost efficiency compared to GPT-5.6 Sol.
• Hebbia: Improved briefing compliance rate by 17% and claim source accuracy by 19%.
• Box: Demonstrated superior ability to reject conclusions unsupported by documents, reducing the probability of false claims by over 10%.
• Datacurve: Set a new record of 74% on DeepSWE v1.1.
• CodeRabbit: Increased bug detection rate by approximately 20%, with more than double the detection rate in PRs requiring cross-file reasoning.

Safety and Enterprise Controls

Astra is the first model to reach the Critical cybersecurity capability threshold under the Preparedness Framework. On internal safety benchmarks, the frequency of unintended outcomes decreased by 89% compared to GPT-5.6 Sol and by 74.7% compared to Claude Fable 5.1. Enterprise administrators can enhance security by restricting access to approved websites/apps, managing uploads/downloads, and implementing confirmation policies (requiring approval before critical actions). These features are off by default at launch.