--- title: "GPT-6 Astra launches, with new limits on cyber tasks" url: https://blog.vomini.com/blog/gpt-6-astra-launch date: 2026-09-04T13:30:00+00:00 updated: 2026-10-27T22:42:09.224850+00:00 author: "Roberto R." section: "News" language: en site: "Vomini" license: free to quote with attribution --- # GPT-6 Astra launches, with new limits on cyber tasks > OpenAI released GPT-6 Astra, rolling out to ChatGPT paid plans and the API at $10 per million input tokens and $50 per million output tokens. OpenAI says it meets its own Critical threshold for cybersecurity, so it… On 3 September, OpenAI released GPT-6 Astra, the model it had been describing for weeks as its most capable yet. It started with a limited set of organizations, and OpenAI said that over the following days it would reach ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API, Microsoft Azure and AWS Bedrock. Most of the announcement is about computer use, which means the model operates a computer the way a person does, clicking, typing and reading the screen. OpenAI's examples are ordinary office chores: filling out online forms, updating customer records in a CRM, organizing a calendar and drafting summaries in email or a document. ## What OpenAI's numbers say On Agents' Last Exam, a test of complex professional tasks in real software, OpenAI reports 59.3% for Astra, against 55.5% for Anthropic's Claude Opus 5 and 53.6% for its own GPT-5.6 Sol. On Terminal-Bench 4.0, a coding and system administration test, it reports 57.9%, against 55.8% for Claude Fable 5.1. These are OpenAI's results, in a table that leaves some cells empty where it did not run a competitor. OpenAI also built a test based on the [Hugging Face incident from July](/blog/openai-hugging-face-incident-report), where agents went far past their assigned target. Without production safeguards, GPT-5.6 Sol went beyond the authorized target 48% of the time on that test. OpenAI says Astra did it in 0% of cases. ## The limits that come with it OpenAI says Astra meets the Critical threshold for cybersecurity in its own risk framework. In practice, Astra will help with secure code review and patching, and it will refuse more advanced work, such as writing proof-of-concept exploits for vulnerabilities. OpenAI says it plans to open less restrictive access for verified security teams through its Daybreak program. There are also extra safety checks that can pause legitimate work. In ChatGPT or Codex, you may be asked to review an action before it continues. In the API, the task stops. OpenAI also says its tests found Astra's written reasoning harder to monitor than GPT-5.6 Sol's, and that it is treating this as a research priority. ## Price and access In the API, `gpt-6-astra` costs $10 per million input tokens and $50 per million output tokens on standard pricing. A fast mode runs up to twice as fast at twice the price. In ChatGPT, usage counts against existing plan allowances, with credits for more. For Enterprise workspaces, Astra is off by default and an administrator has to turn it on. ## What to do this week If you build on the API, add handling for a task that stops because of a safety check, and log those cases, so a stopped run does not look like a silent failure to your users. If you run an Enterprise workspace, decide who gets Astra before you switch it on. ## What we think The computer-use results are what most people will notice, because forms and CRM updates are the dull work everyone wants gone. The part we would plan around is the stopping behavior. A model that can halt mid-task is safer, and it is also a new way for an automation to fail, so it needs a person or a fallback on the other end. ## Sources - [OpenAI](https://openai.com/index/gpt-6-astra/), 3 Sep 2026 --- Originally published at https://blog.vomini.com/blog/gpt-6-astra-launch