OpenAI Launches GPT-6 Astra for Trusted Access Program Participants
OpenAI has introduced GPT-6 Astra, initially available to participants in its Trusted Access Program, with a broader rollout expected soon.
By The Register
OpenAI has unveiled GPT-6 Astra, granting initial access to organisations participating in its Trusted Access Program. This model was announced on 3 September 2026, as reported by The Register.
The launch follows a brief pause in training while OpenAI assessed its AI security measures. GPT-6 Astra is positioned as the top-tier model in OpenAI's current offerings.
The rollout is staged, with wider access anticipated for ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS. The report does not specify any UK venues, businesses, councils, or local residents, so no Cheshire-specific locations can be confirmed from the available material.
OpenAI describes GPT-6 Astra as its most intelligent model yet, featuring advanced capabilities in computer use, browsing, software engineering, science, and professional work. The model arrives shortly after rival Anthropic launched Fable 5.1, both priced at $10 per million input tokens and $50 per million output tokens.
Astra is noted as the first model to achieve the "Critical" level for cybersecurity capabilities under OpenAI's Preparedness Framework. However, some researchers have critiqued the framework, suggesting it may enable the deployment of more dangerous capabilities.
OpenAI claims that GPT-6 Astra represents a significant advancement in cyber capabilities, stating that it can identify previously unknown security flaws and develop new exploitation methods across well-protected systems without human guidance.
The company asserts that Astra is more robust than previous models, with improved alignment, monitoring, and safety in high-risk scenarios. OpenAI has implemented a new evaluation informed by past incidents, indicating that Astra does not exceed its intended scope in challenging situations.
In terms of performance, Astra scored 2 percent on OpenAI's internal hallucination benchmark, a decrease from 9.4 percent for its predecessor, GPT-5.6 Sol. Additionally, Astra is reported to be three times less likely to misrepresent its capabilities compared to GPT-5.6 Sol.
Greg Kamradt, president of the ARC Prize Foundation, noted that GPT-6 Astra surpasses the human baseline in action efficiency on the ARC-AGI-3 test, using fewer actions than the median tested human on 96 percent of levels.
However, the operational cost of Astra is significantly higher, at approximately $360 per game in GPU-forged tokens, compared to the minimal cost of running a human brain.
Astra also brings enhancements to the Codex harness, improving task completion rates and incorporating a new method for context maintenance, allowing it to retain prior session text for future tasks. This feature is currently experimental and will become the default in the coming weeks.