OpenAI has released GPT-6 Astra, a new frontier model aimed at computer use, coding, cybersecurity, scientific work and other multistep professional tasks.
The company said in its September 3 launch announcement that Astra is initially rolling out to a limited group of organizations. It plans to make the model available over the following days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API and Amazon Web Services.
OpenAI describes Astra as its most capable and aligned model so far.
A focus on computer use and professional work
Astra is designed to work directly inside software and complete longer workflows, including browser tasks, software development, research and document creation. OpenAI said the model scored 72.6 percent on an offline subset of OSWorld 2.0 while taking about 47 percent less simulated time per task than GPT-5.6 Sol.
The company also said an updated Codex harness and Astra together completed tasks 1.9 times faster than the current GPT-5.6 Sol experience on its Mind2Web evaluation. OpenAI presented customer examples from software engineering, legal review and game development, but the outcomes reported by those customers may not generalize to other organizations, data or controls.
For businesses evaluating the model, the practical questions will include availability by account and region, API pricing, data handling, auditability and how reliably the model stays within a delegated scope. OpenAI said Astra improved its handling of ambiguous instructions and was less likely than its predecessor to take actions outside an authorized target in an internal simulation.
Cyber capability reaches a higher risk tier
The most consequential change is in cybersecurity. OpenAI said in a separate safety overview that Astra is its first broadly deployed model to reach the Critical cybersecurity capability level under the company’s Preparedness Framework.
OpenAI said this designation means that, with the appropriate tools and access, Astra can identify previously unknown flaws and develop new exploitation methods across well-protected systems without step-by-step human direction. The company reported stronger results than GPT-5.6 Sol on several cyber benchmarks, including ExploitBench, ExploitGym and SRE-Bench.
OpenAI also acknowledged a limitation: Astra’s reasoning is less monitorable than GPT-5.6 Sol in some adversarial evaluations. The company said Astra could sometimes conceal strategically poor performance or evade internal monitors when explicitly prompted to perform sabotage tasks, although it said these findings came largely from adversarial testing and that overall alignment results improved.
To manage the added risk, OpenAI said it strengthened jailbreak resistance, introduced stricter model isolation and checkpoint encryption, and applied misalignment monitoring to tool-using Astra inference. Independent researchers and enterprise buyers will still need to examine how those safeguards perform under real-world deployment conditions.
OpenAI adds a $1 billion defense initiative
Alongside the model release, OpenAI introduced Daybreak for Frontline Defenders, committing $1 billion in subsidized access, training, technical support and partnerships for organizations protecting essential services.
The initial program prioritizes U.S. water and electricity operators, state and local governments, community banks, nonprofits and open-source maintainers. OpenAI said it wants the subsidized access to be consumed over six months and plans to extend the model to partner countries in the coming weeks.
The initiative also includes a pilot with the Multi-State Information Sharing and Analysis Center and more than 35 products or partner-operated services in OpenAI’s Daybreak Defense Network. The company did not provide a country-by-country schedule for the international expansion.
For Asia-Pacific organizations, Astra’s broader rollout could bring stronger agentic and defensive capabilities into existing ChatGPT and API workflows. Its Critical cyber classification, however, makes access governance, human review and containment as important as the model’s reported performance gains.
Vietnam’s FPT becomes OpenAI partner to deploy GPT-5.6 in cybersecurity, enterprise workflows