OpenAI Debuts GPT-6 Astra via 'Trusted Access Program'
AI-generated image · US National Wire
The company's latest model offers advanced cybersecurity and coding capabilities, but initial access is restricted to a select group before a wider rollout.
OpenAI has released GPT-6 Astra, which the company describes as its most intelligent model to date, as The Register first reported. The model is designed for high-level performance in professional work, science, software engineering, browsing, and computer use. However, the rollout is being slow-walked; initial access is granted only to participants in OpenAI's "Trusted Access Program."
According to reporting from The Register, OpenAI intends to expand access to Business, Enterprise, Pro, and Plus subscribers in a few days, provided there are no further security incidents similar to the Hugging Face hack.
Astra is the first OpenAI model to be classified at the "Critical" level for cybersecurity capabilities under the company's Preparedness Framework. OpenAI states that Astra can identify previously unknown security flaws and develop exploitation methods for well-protected systems without human guidance for every step. To mitigate risks, OpenAI claims Astra is more robust and better aligned than previous versions. The company notes that in evaluations regarding whether a model would exceed its intended scope during difficult tasks, GPT-6 Astra had a 0 percent failure rate, compared to 48 percent for GPT-5.6 Sol when production safeguards were absent.
Other performance metrics cited by OpenAI include a reduction in hallucinations; Astra scored 2 percent on an internal benchmark, down from 9.4 percent for GPT-5.6 Sol. Additionally, OpenAI claims Astra is three times less likely to misrepresent its own capabilities than its predecessor, GPT-5.6 Sol.
In terms of external benchmarks, Greg Kamradt, president of the ARC Prize Foundation, stated in a blog post that Astra surpassed the human baseline in action efficiency on ARC-AGI-3, using fewer actions than the median human on 96 percent of levels. On version two of that test, Astra scored 95 percent, while Anthropic's Fable 5.1 scored 90 percent.
However, data from Artificial Analysis suggests Astra trails some competitors in other areas. Astra's Intelligent Index score is 61, which matches GPT-5.6 Sol but is five points lower than Claude Fable 5.1 and also behind Meta Muse Spark 1.3. In coding agent competency, Astra scored 67—on par with Muse Spark 1.3, Fable 5, and Opus 5—but was outperformed by Fable 5.1, which scored 70.
Pricing for GPT-6 Astra is set at $10 per million input tokens and $50 per million output tokens, matching the price of Anthropic's Fable 5.1. Despite the similar token pricing, Artificial Analysis reports that Fable 5.1 is more expensive on a per-task basis at $9.18, compared to $4.72 for Astra.
OpenAI also introduced improvements to the Codex harness, claiming task completion is 1.9x faster. A new experimental feature allows Astra to maintain context across windows by keeping notes and making earlier context windows searchable, a function currently enabled via the Codex config.toml file.