AI Can Use Your Computer Now

Astra feels like the computer-use moment. Five-hour Premiere sessions. Five-day game builds. Blender scenes from a single photo. 99.9% on ARC-AGI-3.

AI Can Use Your Computer Now

Astra will likely be to the rest of computer work what Opus 4.5 was to coding: a capability threshold being passed that leads to a profound reshaping of day-to-day work and large corresponding productivity gains, but for a much larger share of the labor force - Theo Jaffee

This moment has felt like a long time coming. For anyone that has been using frontier models this summer, the trajectory has been clear. Even just a few months ago, the best models would frequently get stuck or take forever on the simplest computer use tasks. With Sol and Fable, things got faster. Now, with Astra, the landscape has shifted dramatically again.

OpenAI’s announcement and countless examples online showcase just how incredible Astra’s computer use is. Navigating websites is a piece of cake. Astra can effortlessly use any computer program you throw at it, and with enough time can navigate and master just about anything.

We became familiar this year with Claude and ChatGPT being more than capable of controlling 3D asset tools like Blender, Unreal, and Unity. With the right set of tools and MCP servers, things that were formerly considered extremely difficult became possible.

Astra is different. It is the first flagship model built around the assumption that software interfaces are machine-operable environments. It doesn’t need every application to expose a clean API. It just treats the screen itself as a programmable interface.

You can look at benchmarks. Astra smashes all of them. It almost doesn’t feel like it makes any sense to focus on the numbers. I think the examples people have posted online from early exposure to the tech are way more interesting.

Let’s dig into how Astra is changing how we think about computer use, and what this indicates about where the tech is heading.

Examples of Astra’s Computer Use

Ethan Mollick’s Personal Wiki

Mollick gave Astra access to a ton of his personal info, GPT-6 downloaded what it needed, constructed a new personal wiki, and designed a whole system for him for personalized briefings. Yes, it sounds a little scary, but this example is a very clear example of where the tech is obviously heading.

Dan Shipper’s Premiere Workflow

Dan and his team had Astra spend five hours operating Adobe Premiere to create a cut of Every’s Fable 5.1 review video. It worked, and the team was impressed. Not kind-of worked. It worked well enough to use. Work like this via computer use was previously impossible.

Claire Vo’s One-Shot Mac Control Tool

Claire had access to Astra for a while, and posted a number of tests, including browser and computer use against her CRM, building various Mac applications, and Blender and 3D asset creations. All of them are impressive on their own, but I was most impressed by Astra’s ability to quickly build a program to control her Mac using hand gestures. She had Astra build the tool while she was waiting for a YouTube video to upload. Enough said.

Tom Krcha’s Blender House Construction

Tom gave Astra a photograph of a house and asked it to build a scene in Blender, including interior assets. The Blender scene was then moved into a walkable Unreal Engine 5 environment. The video says more than any description ever could. One flat image became an editable 3D environment.

This one blew me away. The level of detail is absolutely wild, and it certainly doesn’t look or feel vibecoded.

Matthew Berman’s Five-Day SimCity Clone

Matthew had Astra build a full SimCity-style game using a goal-loop. The model worked for five days straight, creating individual assets and implementing systems for roads, traffic, etc (everything you might expect in a game like SimCity).

The level of detail is staggering. Sure, the model had a full workweek, but the sheer amount of detail is just mind boggling.

The pattern here for the examples I’ve shown is pretty clear. A high level of detail, high accuracy, clear autonomy, effortless computer use, and minimal steering necessary to keep the agents on track.

ARC-AGI-3

ARC-AGI-3 tests how well agents learn as they solve unfamiliar interactive tasks. GPT‑6 Astra saturates the eval, scoring 99.9%.

For context, ARC-AGI is a benchmark used to measure fluid intelligence. It almost seems like OpenAI is making these scores up, but these are legit.

“On ARC-AGI-3, Astra surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark. Not only is this the best model we’ve ever tested, but it also represents a meaningful step change in frontier-model performance - not only in its ability to navigate and solve novel environments, but also in how efficiently it learns to do so.” Greg Kamradt, ARC Prize Foundation

ARC-AGI-4 is scheduled to be released in early 2027, and it seems like by the time the benchmark is released whatever models that are released following Astra will probably ace it, too. ARC-AGI-4 is supposed to measure open-ended invention and autonomous capability gaps.

Greg Brockman, OpenAI’s president, is quoted as saying the following:

“If we fast forward a couple years, and we look back and say when was it really that AGI was created, I think it’s going to be about this time, and I think it might be about this model.”

Astra also scores incredibly well on Agents’ Last Exam, a 59.3% compared with 55.5% for Claude Opus 5.

It seems like at the rate we’re heading, if this truly is the last exam, we may run out of evals for benchmarking by the end of 2027. Agents’ Last Exam is supposed to be extremely difficult. It covers most major fields of professional work performed on a computer.

The demo that may have blown my mind the most was the 15-second condensed playback of Astra performing a PCB layout in KiCad.

A Humbling Human Moment

I love technology, and seeing incredible progress like this is always a humbling experience. Although the trajectory has felt clear for a long time, seeing it finally starting to happen so clearly and so quickly is certainly something else. I remember when Manus first went viral. It was a very visceral ‘woah, I can see where this is heading’. We’re there now. It happened so fast.

If I had to guess, we’ll see increasingly insane Computer Use models throughout the rest of this year, moving into next year. They’ll get smaller, more performant, and eventually there will be versions of Qwen that run locally on your machine that you can use for free. Frontier models like the successors to Astra will zip around your machine so fast you won’t even be able to follow them. Using a keyboard will become increasingly irrelevant.

Computer-based work is going to change. Astra marks the moment clearly. What happens next is for humanity to decide.

- Chris

Photo 1. Deep-space cover art: a luminous spiral galaxy on a black starfield with the word Astra centered in thin white sans-serif type over the galactic core.