OpenAI has launched GPT-6 Astra, which it’s calling the world’s most intelligent and aligned model. The company says Astra brings together years of research and big bets across pre-training, reinforcement learning, and alignment, and that it is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work.
Technical Capabilities & Benchmark Highlights
On native computer use and workflow automation, Astra interacts directly with digital environments by browsing websites, filling out CRM records, managing calendars, and running frontend QA checks. On the OSWorld 2.0 latency simulation, Astra scored 72.6% in roughly 40 minutes per task, outperforming GPT-5.6 Sol (65.7% in 75 minutes) while completing tasks 47% faster.
Regarding context limits, the model features a 1,050,000-token input context window and supports up to 128,000 output tokens, allowing it to process massive repositories, long documentation sets, and complex codebases without breaking context.
Designed as OpenAI’s primary software engineering model, Astra reached 57.9% on Terminal-Bench 4.0 for advanced coding and system engineering. It retains long-horizon conversation context within Codex, allowing it to search past interactions asynchronously to solve complex infrastructure bugs.
For scientific and mathematical reasoning on graduate-level evaluations, Astra recorded 96% on GPQA Diamond, 98% on FrontierMath Tier 4, and 99.9% on ARC-AGI-3.
In cybersecurity capabilities, Astra achieved a 100% score on ExploitBench (without production safeguards) and 42.4% on ExploitGym for vulnerability analysis and defense patching. Production versions include strict guardrails that refuse automated exploit creation.
Benchmark Comparison
| Evaluation Metric |
GPT-6 Astra |
GPT-5.6 Sol |
Claude Fable 5.1 |
| Agentic Coding (Terminal-Bench 4.0) |
57.9% |
37.3% |
55.8% |
| Computer Use (OSWorld 2.0) |
72.6% (~40 mins) |
65.7% (~75 mins) |
— |
| Cybersecurity Vulnerability (ExploitBench) |
100% |
78.5% |
— |
| Scientific Reasoning (GPQA Diamond) |
96% |
— |
— |
API Pricing and Deployment Rollout
Standard API pricing for Astra is set at $10 per million input tokens and $50 per million output tokens. The model is deployable across the OpenAI API, OpenRouter, Microsoft Azure, and AWS Bedrock.
OpenAI is rolling out access across ChatGPT Plus, Pro, Business, and Enterprise accounts. Workspace access is toggled off by default for enterprise administrators to allow security auditing, and eligible enterprise API customers can enable Zero Data Retention (ZDR) policies upon deployment.
GPT-6 Astra marks an operational pivot for OpenAI. Rather than simply retrieving answers or writing code snippets, the model is engineered to execute long-horizon, multi-step workflows across terminal windows, local applications, and enterprise browsers with lower token consumption.
