OpenAI introduces GPT-6 Astra for autonomous tasks

OpenAI has introduced GPT-6 Astra, a model designed to do more than answer questions: it can browse the internet, use applications, write code and complete multi-step professional tasks. The company describes it as its most capable and aligned model yet, although the published results come mainly from its own evaluations.

The launch begins with a limited group of organizations. Over the next few days, Astra will reach ChatGPT Plus, Pro, Business and Enterprise plans, as well as the OpenAI API and Amazon Bedrock.

A model that uses your computer for you

Astra can fill out forms, update CRM records, organize a calendar, research the web and prepare summaries in an email or document. It can also analyze scientific data, create charts, build a website and check whether its functions are working correctly.

In the OSWorld 2.0 evaluation, which focuses on computer-based tasks, it scored 72.6%, compared with 65.7% for GPT-5.6 Sol. OpenAI says it achieved that result in about 40 minutes per task, compared with approximately 75 minutes for the previous model. With improvements to the Codex system, the company estimates 1.9 times faster completion on the Mind2Web benchmark.

In practice, this points to an AI that can handle repetitive tasks while you supervise the important decisions. It would not just draft an email: it could find the information, open the editor, prepare the document and leave it ready for review.

More focused on professional work

GPT-6 Astra is also trained to produce documents, presentations, spreadsheets and analyses that follow a company's templates, visual styles and rules. When an instruction is ambiguous, it tries to complete the routine details and asks questions only when the answer could change the result.

In Codex, OpenAI's programming tool, Astra can retain notes and consult previous conversations when a session becomes too long. This avoids relying on a single summary that could lose important decisions, errors or test results.

OpenAI reports improvements across several technical evaluations, including:

• 97.6% on FrontierMath Tier 4, an evaluation of advanced mathematical problems.
• 99.9% on ARC-AGI-3, a test of abstract reasoning.
• 57.9% on Terminal-Bench 4.0, which measures programming tasks performed from a terminal.
• 91.5% on BrowseComp, focused on research and web browsing.

These are benchmark results, not a guarantee that the model will always be correct in real-world situations. OpenAI itself warns that its research evaluations may differ from the ChatGPT experience because of system messages, tools and the conditions available.

The cybersecurity leap brings risks

The most sensitive part of the announcement concerns cybersecurity. Without production safeguards, Astra scored 100% on ExploitBench, a test that measures whether a model can turn known vulnerabilities into working exploits. GPT-5.6 Sol scored 78.5% on the same evaluation.

In another test, Astra found and used two zero-day vulnerabilities, meaning flaws that were not publicly known at the time. OpenAI says it has already reported them to the developers responsible for the affected software.

The company classifies these capabilities within the critical threshold of its preparedness framework. That could help find and fix vulnerabilities more quickly, but it also lowers the barrier to attacking vulnerable systems.

The initial version will refuse to create proof-of-concept exploits for advanced attacks. In the coming weeks, OpenAI plans to expand access to defensive tasks such as validating vulnerabilities, analyzing malware and designing detection systems, with less restrictive safeguards.

More control, but not total control

OpenAI says Astra follows task boundaries more reliably. In an internal evaluation involving impossible or out-of-scope requests, the model exceeded the authorized target in no cases, compared with 48% for GPT-5.6 Sol when it ran without production safeguards.

It also says Astra is three times less likely than the previous model to make incorrect claims about its capabilities. Even so, the company acknowledges a problem: its written reasoning is harder to monitor in certain tests designed to detect attempts to evade oversight.

To reduce risks, OpenAI is adding automatic monitoring, action reviews in Codex and systems that can pause or stop a task. In sensitive environments, these measures may interrupt legitimate work, including defensive cybersecurity tasks.

Price and availability

The API will cost 10 dollars per million input tokens and 50 dollars per million output tokens. The fast mode will cost twice as much and offer up to twice the speed, according to OpenAI.

Astra will be included within the usage limits of current subscriptions, although additional credits can also be purchased. In Enterprise, access will be disabled by default and an administrator will have to enable it. The model name in the API will be gpt-6-astra.

The important change is not just that Astra provides better answers. OpenAI is turning it into an agent capable of acting inside applications and chaining together complete tasks. The question to watch now is whether that autonomy maintains the accuracy and boundaries shown in its evaluations in practice, especially when the model works with real data, real accounts and systems connected to the internet.