- Published on
A 1.05M-Context Computer-Use Model Gated Behind a ‘Critical’ Cyber Threshold
- Authors
- Name
- DeepHorus
OpenAI has announced the release of GPT-6 Astra, its most advanced artificial intelligence model, which it describes as a highly intelligent and aligned computer-use system. Unlike traditional chatbots, Astra is designed to operate software much like a human would, handling tasks across browsers, spreadsheets, desktop applications, and terminals to complete multi-step jobs directly.
The model represents a significant step forward in AI capability, particularly in its ability to manage complex workflows. A key innovation is its new context-handling system, which maintains notes across context windows and can search back into earlier messages and tool outputs, overcoming a limitation of previous models that often lost critical details during long interactions.
On the technical front, Astra boasts a massive 1,050,000-token context window and a maximum output of 128,000 tokens. It accepts text and image inputs but only generates text outputs. For complex reasoning, it introduces new effort levels, 'xhigh' and 'max', and supports a wide array of tools including computer use, hosted shells, and patch applications.
Performance benchmarks show notable achievements. OpenAI reports that Astra scored 72.6% on the OSWorld V2-Offline test, surpassing its previous GPT-5.6 Sol model. It achieved a 99.9% score on the ARC-AGI-3 benchmark, though this result reflects the combined performance of the model and its agent system.
However, some analysts note that in coding-specific benchmarks, Astra's lead is marginal. For instance, on the DeepSWE v1.1 test, its score of 74.1% is only slightly ahead of competitors like Meta's Muse Spark and Anthropic's Claude Opus 5, indicating that other models remain highly competitive in that domain.
Astra's most consequential feature may be its cyber capabilities, as it is the first OpenAI model designated as meeting the 'Critical' threshold in the company's Preparedness Framework. In testing, it developed exploits for hardened systems and discovered two previously unknown vulnerabilities, prompting a restricted access policy.
Access to Astra is currently limited to organizations in OpenAI's Trusted Access and Daybreak programs. Standard API users will face automatic blocks on advanced cybersecurity tasks, and may experience slowdowns or pauses even during unrelated work. The model costs $10 per million input tokens and $50 per million output tokens.
Looking ahead, Astra's restricted deployment sets a precedent for how advanced AI with potential dual-use capabilities will be managed. The focus on direct computer operation over chat, combined with enhanced memory and reasoning, points toward a future where AI agents could autonomously manage complex digital tasks, albeit within carefully controlled safety boundaries.
