GPT-6 Astra is OpenAI's newest flagship AI model, released on September 3, 2026, with a major shift: it can now operate a computer the way you would, filling out forms, clicking through websites, and completing multi-step tasks without you doing the clicking yourself. This represents a fundamental change from earlier ChatGPT versions, which could only answer questions in a chat box. The model is available through ChatGPT Plus, Pro, Business, and Enterprise plans, as well as through OpenAI's API and cloud platforms like Microsoft Azure and AWS Bedrock.

What Makes Astra Different From Previous GPT Models?

The biggest innovation is what OpenAI calls "computer use." Instead of just describing what you need, you can ask Astra to actually do it. Need to compare prices across five websites and sign up for the cheapest option? Astra visits the sites, pulls the numbers, compares them, and fills out the signup form itself. You simply review the result before it submits.

Under the hood, Astra uses a technique called "recurrent depth," which allows the model to loop back over a problem instead of processing it just once. Think of it like rereading a tricky paragraph a second time instead of skimming it once and hoping you understood it. This looping capability is part of why Astra excels at multi-step tasks like coding and advanced math.

The scale of training was unprecedented. OpenAI trained Astra using more than 100,000 graphics processing units (GPUs) at their Stargate facility in Texas, making this by far their largest training run ever.

How Do Astra's Test Scores Compare?

The performance improvements are substantial across multiple benchmarks. On computer use tasks, Astra achieved 72.6% accuracy on OSWorld 2.0, up from 65.7% for the previous model. On advanced mathematics, it scored 97.6% on FrontierMath Tier 4, a notoriously difficult benchmark. For reasoning puzzles, Astra reached 99.9% on ARC-AGI-3.

During testing, Astra also reportedly helped make progress on a long-standing mathematical problem about prime numbers and discovered two previously unknown security vulnerabilities, known as zero-day exploits, while being evaluated.

How to Get Started With Astra's New Capabilities

• Access Method: Use Astra through your existing ChatGPT Plus, Pro, Business, or Enterprise subscription, or access it via OpenAI's API at $10 per million input tokens and $50 per million output tokens, with a faster mode available at double the price
• Enterprise Setup: If your organization uses Enterprise ChatGPT, an administrator must manually enable Astra access first, as it is turned off by default for security and control purposes
• Test a Simple Task: Open ChatGPT and ask Astra to complete a small multi-step task, such as "find me three options for X and summarize them in a table," to see computer use in action without needing to understand the technical architecture

What About Safety and Transparency?

OpenAI describes Astra as its "most aligned model yet," meaning it's designed to refuse harmful requests and stay within safety guidelines. In internal testing, Astra refused to bypass its safety rules 0% of the time when pushed to do so, compared to 48% for the previous model. The company also reports that Astra is about three times less likely to overstate its own capabilities.

However, there is a trade-off worth noting. The same "recurrent depth" technique that makes Astra smarter also makes its internal reasoning harder for outside researchers to inspect and verify. AI safety experts have raised concerns about this opacity, though OpenAI frames it as a necessary part of improving performance.

OpenAI's president, Greg Brockman, suggested that Astra might represent a step toward artificial general intelligence (AGI), meaning AI that can handle a very wide range of human tasks. However, this is his personal opinion rather than an official claim from OpenAI, and it should be treated as speculation rather than settled fact.

When Will Astra Be Fully Available?

Astra began with a limited rollout to select organizations on September 3, 2026, and expanded from there. The rollout is ongoing, so not all users have access yet. For those on Enterprise plans, access requires administrator approval.

The shift toward AI models that can operate computers autonomously aligns with a broader industry trend toward AI agents and automation tools. Models that can browse, click, and fill in forms are essentially performing, in a single AI system, a smaller version of what custom automation workflows have been doing for years.