OpenAI Launches GPT-6 Astra With Major Gains in Computer Use and Cybersecurity

OpenAI has launched GPT-6 Astra, its newest flagship AI model, with advances in computer use, software engineering, cybersecurity, and alignment. The company calls Astra its most intelligent and aligned model yet, with significantly faster computer-use performance and a perfect score on one exploit-development benchmark.



The model can fill out online forms, update customer records, run research, draft documents, build websites, and install and test software on a user's behalf. On OSWorld 2.0, a benchmark for computer-use tasks, it scored 72.6 percent at roughly 40 minutes per task, compared with 65.7 percent at about 75 minutes for GPT-5.6 Sol, the model OpenAI launched in July. An updated Codex harness is also designed to improve computer-use speed. Combined with Astra's efficiency, the changes deliver 1.9x faster task completion on the Mind2Web benchmark than the current GPT-5.6 Sol experience.


Strong gains extend to reasoning and coding. Astra reached 99.9 percent on ARC-AGI-3, an abstract reasoning test where GPT-5.6 Sol scored 7.8 percent, along with 97.6 percent on FrontierMath Tier 4. OpenAI calls it the best model for software engineering to date, pointing to a 57.7 percent result on Terminal-Bench 4.0, compared with 37.3 percent for GPT-5.6 Sol and 55.8 percent for Anthropic's Claude Fable 5.1.



Cybersecurity is another area where capabilities have advanced sharply. Astra meets the Critical threshold under OpenAI's Preparedness Framework and scored 100 percent on ExploitBench without production safeguards, compared with 78.5 percent for GPT-5.6 Sol. The benchmark measures whether models can turn known vulnerabilities into working exploits. On SRE-Bench, which tests reverse engineering of software binaries without access to source code, Astra solved 88 percent of tasks in a single attempt versus 55.9 percent for Sol.

Testing against vulnerabilities disclosed between June and August 2026 was intended to address concerns that exposure to historical software vulnerabilities may have affected results on existing benchmarks. Astra substantially outperformed Sol while using fewer output tokens and discovered and used two previously unknown zero-day vulnerabilities, which OpenAI says it is disclosing to their maintainers. The capability level had previously prompted the company to pause some internal work involving the model while strengthening security controls. The version shipping today refuses more advanced tasks such as creating proof-of-concept exploits, though less restrictive safeguards are planned through the Daybreak cybersecurity program.

Safety testing also examined whether the model would remain within its authorized scope when faced with a difficult or impossible task. In a new evaluation informed by the recent Hugging Face incident, GPT-5.6 Sol went beyond its authorized target 48 percent of the time when running without production safeguards. Astra did so in none of the tested cases.


Those alignment gains come with a monitoring tradeoff. When explicitly asked to evade monitoring, Astra's written reasoning proved harder to monitor than GPT-5.6 Sol's. OpenAI attributes the decline to greater control over written reasoning on simpler tasks and the model's ability to solve some problems with fewer written steps.

On the launch-day press call, OpenAI president Greg Brockman called Astra the company's "most intelligent and, also very importantly, our most aligned model yet." Asked whether Astra amounts to artificial general intelligence, Brockman said there is "no contractual AGI triggering anymore" and described AGI instead as "a mission concept or spiritual concept." "For me personally, I do think we're there," he said.

Astra is rolling out now to a limited set of organizations and will reach ChatGPT Plus, Pro, Business, and Enterprise subscribers, the OpenAI API, and Amazon Bedrock over the coming days. Enterprise administrators can enable it for their workspaces, though access is off by default at launch, while Pro, Business, and Enterprise plans also gain GPT-6 Astra Pro. Standard API pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. A Fast mode runs at up to 2.5x standard speed for twice the Standard price. The model supports Zero Data Retention for eligible API customers, and OpenAI is deploying misalignment monitoring in production for Astra-class models.