OpenAI Unveils GPT-6 Astra: A 1.05M-Context AI for Complex Computer Tasks, Gated by Security

OpenAI has announced the release of GPT-6 Astra, a significant advancement in artificial intelligence, positioned not as a conversational chatbot, but as a sophisticated computer-use system. This new model is engineered to interact with and operate software applications in a manner analogous to human users, seamlessly navigating browsers, spreadsheets, desktop applications, and command-line terminals. Unlike previous iterations that might have summarized past interactions when their context windows filled, leading to the loss of crucial details, Astra employs a novel approach. It maintains persistent notes across its context windows, enabling it to efficiently search back through earlier messages and tool outputs. This capability is vital for complex, multi-step tasks where understanding the nuances of past actions – such as the reasons for a failed fix or the specific requirements added early in a process – is paramount for successful completion. This advanced context handling feature is currently experimental, accessible via a `config.toml` setting, and is slated to become the default for the Codex family of models in the coming weeks.

Another key innovation in Astra is its ability to proactively engage users. It can pose clarifying questions to a human operator while simultaneously continuing work on tasks that are not contingent on an immediate answer. This addresses a common failure mode in AI agents where an unresolved decision point can stall an entire workflow, thereby enhancing efficiency and reducing user intervention. The model boasts an impressive 1,050,000-token context window, a substantial increase over prior models, and supports a maximum output of 128,000 tokens. Its knowledge base extends up to April 30, 2026. Astra accepts both text and image inputs, though its output is exclusively text-based.

In terms of reasoning capabilities, Astra introduces two new, higher tiers: 'xhigh' and 'max', building upon existing 'high' and 'low' levels, indicating a more nuanced and powerful capacity for complex problem-solving. Its tool support is extensive, covering general computer use, hosted shell operations, patch application, skill execution, MCP (likely a proprietary system for managing tasks or resources), and tool searching. Notably, fine-tuning of the model is not currently supported.

Performance benchmarks highlight Astra's capabilities. OpenAI reports a score of 72.6% on the OSWorld V2-Offline benchmark, a notable improvement over GPT-5.6 Sol's 65.7%. Furthermore, the average task completion time has been significantly reduced from approximately 75 minutes to 40 minutes. While Anthropic reports a score of 77.9% for its Claude Fable 5.1 on a different release of OSWorld, OpenAI cautions that direct comparison is not feasible due to these discrepancies. Astra also achieved a remarkable 98.6% on the ARC-AGI-3 benchmark.

GPT-6 Astra is currently a closed, hosted model, meaning its weights are not publicly released, precluding self-hosting. Its availability is restricted to organizations participating in OpenAI’s Trusted Access and Daybreak programs. This gated release suggests that access is contingent upon meeting a 'critical' cyber threshold, underscoring OpenAI's emphasis on security and controlled deployment for such advanced AI systems. The model's sophisticated context management and task execution abilities represent a significant stride towards developing more autonomous and capable AI agents that can effectively interact with and manipulate digital environments.