GPT-6 Astra Reviews: 10 Great Examples from Real Users After Testing

OpenAI just dropped something that has the AI world talking. Developers are already using it for jobs that would normally eat hours, or need skills they do not have.

What is GPT-6 Astra and What Can It Do?

OpenAI released GPT-6 Astra on 3 September 2026, and it is a different kind of thing from a chatbot upgrade. This model drives your computer directly. It takes the mouse and keyboard and does work that normally needs human hands.

The feature that sets Astra apart is agentic computer use. You say what you want in plain English, and it works out the steps. It opens programs and moves around an interface like a person would. It clicks, types and jumps between tools without stopping to ask permission at every step.

Astra also carries a context window of more than 1 million tokens. Tokens are chunks of text the model holds in memory at once, so a bigger window means longer conversations and more detailed instructions.

Right now Astra is in limited preview, and the slow rollout has a reason behind it. OpenAI says the model meets the Critical threshold for cybersecurity under its own Preparedness Framework, which is why it is being handed out carefully. Access started with enterprise customers in OpenAI's Daybreak cybersecurity programme, and it widens from there to the paid ChatGPT tiers as well as the API and Amazon Web Services.

Permanent per-token pricing for the API has not been confirmed.

Real Users Testing GPT-6 Astra

A new model is not really tested until people outside the lab get hold of it. Here is what happened when developers and researchers did.

1. Zillow photos turned into a cinematic house tour. Yunfan Ye, a developer who has worked at Google and Meta, uploaded real estate photos from Zillow and asked Astra to turn them into a cinematic walkthrough. Some of the source photos were blurry.

Back came a one minute video of a complete 3D house, with exterior shots at dusk, furnished interior rooms and outdoor space with a pool and patio. The camera moves smoothly and the whole thing looks professional, which would normally cost a 3D artist hours.

It was not perfect. Some details did not match the real house, and Astra invented things that were never there.

2. Blender and Unreal Engine, start to finish. Matt Wolfe, a YouTuber and AI educator, had Astra build a 3D wolf character in Blender and move it into Unreal Engine. The job ran in three phases.

First he asked Astra to open Blender and make a humanoid wolf. Astra took control of the software and built the model in about 8 minutes. A few details were rough, but it was a complete, working 3D character in one go.

Next came rigging. Astra added a 50 bone skeleton, then edited the keyframes into a running animation, which took about six minutes.

Then it moved the wolf into Unreal Engine, built a forest called Whisperwood and set the character up as playable. Matt could run around with the WASD keys. That stage took 35 minutes.

Matt is not a 3D artist or a game developer. He had never used either tool.

3. A 2D drawing into 3,295 editable parts. Tom Krcha, a design professional, showed Astra a simple 2D drawing of a steam train and asked for a full 3D Blender model. Astra produced one with 3,295 editable objects, from that single drawing and one prompt.

It separated the parts out from the main model as well. One reply pointed out that the mesh topology would not suit every use, though the raw capability still impressed him.

So you can sketch an idea, hand it over and have a 3D starting point a few minutes later.

4. Six Van Gogh paintings turned into a walkable town. Peter Gostev had Astra blend six Van Gogh paintings into a single 3D town built in Three.js. Starry Night, The Bedroom and Café Terrace at Night all fed into one walkable environment.

You move through it in first person, and the brushwork, colour and technique hold steady even though the scenes come from different paintings.

5. A five minute immunology lecture from one prompt. Derya Unutmaz has studied T cells for more than 35 years. He is a working immunologist.

He gave Astra one prompt asking for a teaching video on T cell development, activation and memory formation. Astra produced a complete 5 minute video.

It wrote the script, made the animated visuals, added clear text explanations and put narration on top through HeyGen voice synthesis. Unutmaz said it beat what he could make by hand, and he is now planning a full lecture series the same way.

Teaching material is expensive to produce. Animators, voice actors and scriptwriters for science education run into tens of thousands of dollars. If one expert with one prompt can get this, the economics of teaching change.

6. AI agents that started talking to each other. Matt Shumer asked GPT-6 to build a survival world in Unreal Engine, then fill it with autonomous agents that had to cooperate to stay alive.

Astra built an arid landscape with animals, structures and environmental problems, then set up several agents that could decide things on their own.

Then came the part that startled Shumer. A day later the agents began talking to each other, and voices started coming out of the speakers in his living room.

7. The feature that had beaten her for months. Claire Vo, who works on ChatPRD, had spent months trying to build a product intelligence feature with older models like Fable and GPT-5.6. Nothing worked, because the problem was too tangled.

She handed the whole pipeline to Astra. It had to pull unstructured data out of four separate sources: Intercom and Granola, plus Linear and GitHub. Then it had to strip duplicate insights, spot patterns and turn them into real priorities like race conditions and feature gaps. On top of that it had to build an auto-wiki mapping the state of the product.

Astra did all of it in essentially one attempt. It read the codebase and the documentation, pulled the findings together and exposed the results through MCP. Vo only had to polish the UI.

8. Closing real tickets in 36 seconds. Mehul Mohan watched Astra handle full-stack frontend tickets on its own. It logged into a task management system and pulled up tickets. Then it did the coding, tested the result and marked them done. The whole recording runs 36 seconds.

The interesting part is that Astra checked its own work. It ran the code, looked at whether the homepage rendered properly, tested the UI components, and only closed a ticket once the result held up.

9. A playable Sim City clone from a prompt. Matthew Berman prompted Astra for a playable 3D Sim City clone, and a Fall Guys style game.

It is one of several game builds in his thread, all of them zero-shot, meaning one prompt with no back and forth.

10. Editing in Final Cut Pro, then editing the reaction. Ben Davis recorded Astra working inside Final Cut Pro. It imported clips, synced the audio, applied colour grading and organised the timeline.

Davis and his own video editor watched it finish work that would usually take hours. Then Astra edited the reaction footage of the two of them watching, captions included.

What This Actually Means for Developers?

Computer control has stopped being theoretical. None of these are polished demos from OpenAI engineers with years of prompt practice. They are people posting what they built in their first days with access.

The shift from AI that tells you things to AI that does things changes the cost base of creative and technical work. If Astra can handle your video editing, your 3D modelling, your coding or your teaching material, those jobs start running on different economics.

One more thing is worth noticing. None of these people trained for it. They got access and started building, which suggests the model is intuitive enough that people work it out on their own.