Back in February, getting Codex to build LEGO models took a custom skill, complicated instructions, and hours of iteration. And it was still not great. I came away thinking it wasn’t something models were good at.
With GPT-6 Astra, I gave it access to BrickLink Studio and asked for the Golden Gate Bridge. About ten minutes later, I had a first version, complete with little cars on the road. I kept refining it afterward.
Experiences like that have made me reconsider how I use Codex: which apps I give it, how much I specify, and how much reasoning a task needs. Here are five things I learned from using Astra.
1. Astra can use all your apps (sometimes, better than you do) 2. Astra does more of the product thinking 3. Astra excels at Blender 4. Astra doesn’t need Max to impress 5. Astra keeps checking its work
1. Astra can use all your apps (sometimes, better than you do)
Let’s go back to that LEGO bridge. I’d already seen Astra work in other apps, so I wanted to see how it would handle BrickLink Studio, the app for designing digital LEGO models.
I gave it a simple request:
Prompt: "Use Studio to build the golden gate bridge in lego"
Here’s that first version:
The first bridge, rendered afterward with studio lighting.
It even added the little cars.
Astra figured out what parts were available in Studio’s bundled library, used Python to assemble the model, then opened it in Studio to inspect and iterate on it.
What surprised me was how naturally it combined code and the app. It’s fast enough that I’m starting to question whether doing the task myself would actually be quicker.
It’s also made me revisit the skills we have built for previous models. When we were having Astra edit videos for the launch, some of our old skills were getting in its way. A more vanilla setup, without the video editing skills, worked better for that task.
If you have a skill built around something an older model couldn’t do, try the task without it and see whether you still need it.
Computer use also lets Astra connect work across apps. It can create something in one place, inspect it somewhere else, and bring the pieces together. Giving it that access has helped it verify its own work and complete more of the task.
So think about which app you’d use for the job and give Astra access to it. That could be Bricklink Studio for LEGO, Blender for 3D, or DaVinci Resolve or a similar editor for video. It doesn’t have to do all of its video work through FFmpeg if there’s a dedicated app that can help.
Tip: Give Astra the apps you’d use for the job and try it without extra skills first. It may need less guidance than you think.
2. Astra does more of the product thinking
For one test I wanted to build Thumbnail Studio to help make creative assets. I was able to quickly ramble through some ideas and reference conversations I’d had with a colleague. Astra figured out a lot of the detailed interactions you’d expect from an image editor in the browser, including ⌘-Z and other familiar shortcuts, layers, and the ability to hide them. I didn’t have to be as prescriptive as I would have been previously.
The prompt I gave it and the first version it built.
Thumbnail Studio after more iteration.
Later, I asked for a Product Feedback mode and an automation to check, rank, and implement the feedback. What surprised me was how thoroughly Astra thought through the feedback interface. I could highlight parts of the app, pin comments to them, set priorities, edit feedback, and mark it as resolved. I hadn’t described any of those interactions in detail.
The prompt I gave it and the feedback mode it built.
I could quickly fire off a task for a capability I wanted, and Astra would work through what the experience should look like. That’s making it much easier for me to build personalized software. I can react to what it builds instead of planning every interaction before I start.
The animations in my WebMCP article and in this article are another example. I liked Nick Baumann’s animation on X, so I gave Astra an Appshot of it, the article I was writing, and access to the Codex app codebase. The Appshot included a screenshot and context from the app.
Codex read the article, understood the 3D demo I was referring to, and recreated a full 3D interface it could animate. It thought through the composition, the camera angles, and the text explaining what was happening. Then it built me a little preview tool where I could play, pause, repeat the animation, and view it from different angles.
I hadn’t asked for that preview tool, but it made it much easier to iterate. I could give it detailed feedback using the Annotations tool in the in-app browser.
The preview tool, the first animation, and the version I published.
In some cases this behavior might back fire for you, if you have very specific things in mind make sure you are prescriptive with your expectations.
I’m also spending less time preparing everything the model needs before I ask it to do something. To make the comparison above, I asked Astra to go through the old conversation and recover an earlier version of the animation, find a screen recording somewhere on my desktop, and get the final animation from the X article.
It found the pieces and combined them into a video without me having to point it to every file. I could describe the result I wanted and let it figure out how to bring everything together.
When I asked for a site to showcase its LEGO models, I wanted people to browse them in 3D and download the build files. Astra also added an interface for uploading more models. It gated that interface with Sign in with ChatGPT so only I could use it as the admin.
I still need to know what I want and give it a direction to get the most out of Astra. For the showcase, I gave it more examples and asked for interactive, rotating planets to get the space theme I had in mind. But I don’t have to be pedantic about every feature. The better the references I give it, the more Astra surprises me with how carefully it thinks through the experience.
Tip: You don’t have to spell out every interaction up front. Give Astra the idea and a few good references, if you want it to act narrowly tell it the requirements, then refine the experience it builds.
3. Astra excels at Blender
Our team had already been impressed by the 3D experiences Astra could build with Three.js or even raw WebGL. But one thing that stood out was how well it could use Blender, both through computer use and the command line. Even when the final result belongs in a browser, I’ve found it useful to let Astra build parts of it in Blender.
To test that, I gave it ten quick photos of a home bar and asked it to rebuild the room in Blender. I didn’t give it any measurements.
Astra was meticulous about the proportions. It found reference objects in the photos, such as the cocktail books, and used them to estimate the scale. It modeled individual objects, checked them separately, placed them in the scene, and kept iterating. It built the room, furniture, shelves, bottles, and glassware. It even nailed the chair in the corner and random wires in the room.
Having the individual objects in Blender let it inspect and refine each part of the scene. I kept iterating on the materials and lighting, and I could open the scene and look around from different angles.
The home bar rebuilt in Blender from ten photos.
Tip: If you are working on a project with 3D ask Astra to use Blender for better results.
4. Astra doesn’t need Max to impress
How far I can get on Light/Low reasoning has been one of the biggest surprises. With every new model generation, the temptation is high to try to use the highest level of reasoning. Especially if that’s what you were using before.
In my Blender experiments, I’ve had results from Astra on Low that I prefer to Sol on Max. Astra is also very thoughtful in verifying its work regardless of reasoning level, making Low reasoning incredibly powerful.
My Blender experiments with Sol on top and Astra below, from Low to Max.
If you’ve been using Sol on high reasoning, don’t assume you need to crank Astra up too. Try something like Medium first, then experiment in both directions and see what works for the task. If you need some inspiration of more comparisons, check out our Model Observatory.
Tip: Start with Light/Low or Medium reasoning. Look at what you get before deciding you need more reasoning.
5. Astra keeps checking its work
Astra tends to check that it actually completed the task. If you are building a game and it might start playtesting it. Ask it to optimize something and it might set up an environment to measure whether the optimization worked. For a video, it might inspect individual frames or even sound waveforms to check what it’s doing.
When I had Astra work on the title image for this article, it used browser use to check the code behind X’s live article page and confirm the header image requirements. It found a 5:2 aspect ratio and a 1200 × 480 cover.
It checked the live page’s code before confirming the header format.
That can be useful, but I need to be clear about how far I want it to go. If you want a prototype you’ll playtest yourself, say that. If you want it to play through the game and fix what it finds, say that too. For a final check, you could ask it to make one pass through the checks you care about, then report what it found and anything still unresolved.
Tip: Tell Astra what the finished result should be, what you want it to verify, and when it should hand the work back.
Try the task you gave up on
I’d spent hours trying to get Codex to build LEGO models and came away thinking it wasn’t very good at it. With Astra, I had a first version of the bridge in about ten minutes.
Give it the apps you’d use for the job and a few good references. Start with Light or Medium reasoning. Tell it what the finished result should be and what you want it to verify, then see how far it gets before adding more guidance.
If you’ve got a task you tried with an older model and gave up on, give it another go with Astra.
