OpenAI Unveils GPT-6 Astra, Claims Breakthrough in AI Alignment and Autonomous Work
OpenAI's GPT-6 Astra launches with claims of superior alignment and autonomous computing, scoring high on benchmarks and safety tests.
OpenAI has unveiled GPT-6 Astra, a new artificial intelligence model it describes as its most intelligent and aligned creation to date. The system is engineered for a range of complex tasks, including computer browsing, software engineering, mathematics, and enterprise workflows.
The company announced that deployment begins immediately for a select group of organisations, with broader access to follow across consumer and enterprise subscription tiers. OpenAI highlighted that Astra integrates years of research in pre-training, reinforcement learning, and alignment, positioning it as state-of-the-art in computer use, cybersecurity, science, and professional work.
On standardised benchmarks, the model recorded a 98 per cent score on FrontierMath Tier 4, a 99.9 per cent mark on ARC-AGI-3, and a 100 per cent completion rate on ExploitBench. OpenAI also noted that the system assisted in resolving open mathematical problems before its commercial release.
A key focus of the announcement was systemic alignment. OpenAI stated it developed new testing protocols, informed by earlier industry events, to ensure the model adheres strictly to operator boundaries. In one evaluation, GPT-6 Astra went beyond its authorised target in 0 per cent of cases, compared to 48 per cent for its predecessor, GPT-5.6 Sol, when production safeguards were removed.
The model also demonstrates autonomous administrative capabilities, such as CRM data entry, website quality assurance, and independent software installation. On the OSWorld 2.0 benchmark, it completed tasks 47 per cent faster than GPT-5.6 Sol, achieving a 72.6 per cent success rate.
GPT-6 Astra will be available through ChatGPT Plus, Pro, Business, and Enterprise plans, as well as via the OpenAI API, Microsoft Azure, and AWS Bedrock. The company confirmed support for Zero Data Retention for eligible API customers and is testing Private Safety Processing to enhance monitoring without compromising privacy. Developer access is priced at USD 10 per million input tokens and USD 50 per million output tokens, with a high-speed option available at double the cost.