OpenAI released GPT-6 Astra yesterday, calling it their “most capable model for business.” The headline feature is computer use. The model can now control computers directly, clicking through interfaces and navigating software on your behalf.

This isn’t the first computer-use model. Anthropic shipped something similar with Claude 3.5 Sonnet in October 2025. But GPT-6 Astra represents OpenAI’s entry into what’s becoming a clear trend: AI models that don’t just advise, they operate.

What’s actually new

Beyond computer control, OpenAI is pitching two other improvements: advanced reasoning and what they’re calling “stronger writing and design judgment.”

The reasoning piece likely builds on the o1 architecture OpenAI released earlier. These models spend more time thinking through problems before answering, trading speed for accuracy on complex tasks.

The design judgment claim is harder to evaluate without testing. OpenAI is suggesting the model can make aesthetic calls, not just functional ones. Whether that means better UI mockups, cleaner prose, or something else entirely isn’t clear from the announcement.

Who this is for

OpenAI is explicitly positioning Astra for business users. The announcement post focuses on work scenarios: automating workflows, handling complex analysis, operating enterprise software.

If you’re already using ChatGPT Enterprise or building on the API, this is worth testing. Computer use could genuinely speed up repetitive tasks, especially the kind that involve clicking through multiple tools.

For individual developers or hobbyists, it’s less clear. OpenAI hasn’t announced pricing yet, and the “for business” framing suggests this won’t be cheap. If you’re just prototyping or doing side projects, GPT-4o or Claude 3.5 Sonnet are probably still the better value.

The computer use question

Here’s the thing about computer-use models: they’re powerful and kind of terrifying. When a model can click buttons and fill out forms on your behalf, you’re trusting it with a lot. One wrong action in the wrong system and you’ve got a real problem.

Anthropic has been cautious about this, releasing computer use in beta with warnings about reliability. OpenAI’s approach seems more aggressive, positioning it as a core feature rather than an experiment.

That confidence might be justified. Or it might mean we’re about to see some spectacular failures in production environments. We’ll find out soon enough.

Also shipped: AI agents get email addresses

While we’re on the subject of AI tools doing things autonomously, Instinct launched email support yesterday. The viral AI assistant can now create and manage email accounts, contact businesses, handle support requests, and generally operate as a digital proxy.

This is the logical extension of AI agents. If they’re going to book appointments and manage subscriptions, they need a way to communicate with systems that still run on email. Instinct is calling it a “professional identity” for your AI.

It’s clever. It’s also unsettling. We’re normalizing the idea of AI agents operating independently in systems designed for humans. That’s fine when it works. When it doesn’t, good luck figuring out what your agent agreed to on your behalf.

The pattern is clear: AI tools are moving from assistants that suggest to agents that act. GPT-6 Astra and Instinct’s email feature are both pushing in that direction. Whether enterprises are ready for that shift is a different question.