Autonomous Computer Use and Professional Workflows

GPT-6 Astra represents a shift toward agentic capabilities, specifically in its ability to interact with computer interfaces to perform multi-step tasks. It achieves a 72.6% score on OSWorld 2.0, completing tasks 47% faster than its predecessor, GPT-5.6 Sol. The model is designed to handle tedious administrative work—such as CRM updates, calendar management, and data analysis—while maintaining higher accuracy in following user-defined templates and visual styles.

Enhanced Software Engineering and Context Management

Astra introduces a new approach to context window management in Codex. Instead of relying on lossy compaction, the model now maintains searchable notes across sessions, allowing it to retrieve specific details about past debugging efforts or component behaviors without needing to re-summarize history. This improves performance in complex refactoring and long-running software engineering tasks.

Cybersecurity and Safety Alignment

OpenAI has positioned Astra as its most aligned model, reporting a 0% failure rate in tests where models were prompted to exceed authorized targets (compared to 48% in previous versions). In cybersecurity, the model demonstrates advanced capability, achieving a 100% score on ExploitBench and discovering two previously unknown zero-day vulnerabilities during internal testing. To mitigate risks, OpenAI is implementing a tiered safety approach: while the model can assist in defensive tasks like secure code review and patching, it currently refuses to generate proof-of-concept exploits. These restrictions are managed via a combination of production safeguards, such as Codex Auto-review, and active misalignment monitoring that can halt unauthorized actions in real-time.

Scientific and Reasoning Benchmarks

Astra shows significant improvements in abstract reasoning and scientific application. It achieved a 99.9% score on ARC-AGI-3 and a 97.6% score on FrontierMath Tier 4. These capabilities enable the model to work within specialized scientific software to inspect data and assist researchers in determining next steps for investigation.