OpenAI Launches GPT-6 Astra, a Model Built for Autonomous Computer and Coding Tasks

OpenAI has unveiled GPT-6 Astra, a large language model engineered to perform self-directed tasks across coding, professional workflows, science, and cybersecurity use cases. The model is becoming available progressively to ChatGPT Plus and Pro users, enterprise clients, and developers via the OpenAI API, Microsoft Azure, and AWS Bedrock.

Computer Interaction & Multi-Step Automation#

Rather than simply generating text responses, Astra is designed to execute workflows by directly manipulating graphical interfaces. The model can autonomously complete a range of practical tasks: filling web forms, managing customer relationship management systems, conducting online research, building websites, processing datasets, deploying software, running tests, and diagnosing visual problems on-screen. On industry benchmarks like OSWorld 2.0—which measures real-world software usability—Astra achieves 72.6%, up from 65.7% on GPT-5.6 Sol.

Enhanced Coding & Long-Context Performance#

Coding represents a major focus area for the release. The model demonstrates strong results on standard coding benchmarks: 57.9% on Terminal-Bench 4.0 and 74.1% on DeepSWE v1.1. A new experimental feature in Codex allows the agent to maintain persistent notes and context across extended sessions, avoiding the loss of information that typically occurs with standard context-window limitations. This enables the model to reference earlier requirements, test outputs, and diagnostic data during lengthy development tasks.

Astra supports up to one million tokens of input context, achieving 96.3% accuracy in retrieval tasks spanning the 512K-to-1M range, unlocking use cases that require sustained reference to large codebases or documentation.

Cybersecurity Capabilities & Risks#

A watershed moment for OpenAI's safety approach: Astra is the first OpenAI model to reach the critical capability tier in its Preparedness Framework for cybersecurity. During internal testing conducted without production safeguards, the model discovered and leveraged two previously undocumented vulnerabilities and devised exploit code targeting hardened browser and operating system environments.

The production release includes guardrails restricting offensive security operations. OpenAI plans to roll out deeper defensive capabilities through its program, positioning the tool for use by security researchers and blue-team defenders.