I Gave GPT-6 Astra Blender Access (The Results Are Insane)
Pat Simmons lets GPT-6 Astra run Blender completely unsupervised for over 54 hours across three builds: a historic house, his own studio, and a robot's website.
GPT-6 Astra can drive Blender through its Python API and MCP bridge for 20+ hours at a stretch without stopping, researching references, pulling in outside assets on its own initiative, and QAing its own geometry with computer-use screenshots to turn a single riffed prompt into a near-photoreal 3D model.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're curious whether AI coding agents can actually operate creative software like Blender unsupervised, not just write code.
You're in real estate, architecture, interior design, or construction and want a cheap way to produce 3D walkthroughs or virtual staging for clients.
You run a website and want a premium, interactive 3D hero element without hiring a 3D artist.
You already pay for a flat-rate AI subscription and want to know what a very long, high-effort agent run can actually accomplish inside it.
SKIP IF…
You need a fast turnaround; each build here runs 20 to 54 hours unattended.
You want a hands-on Blender tutorial; Pat doesn't know Blender himself and is directing an agent, not teaching the software.
You're looking for a free or open-source path; the economics here depend on an existing $200/month ChatGPT Pro subscription.
TL;DR
The full version, fast.
Pat Simmons hands GPT-6 Astra, running at 'extra high' effort in a coding CLI, three unsupervised Blender builds: a pixel-for-pixel recreation of Frank Lloyd Wright's Falling Water, a scan of his own studio from a shaky phone video, and a 3D render of the 1X Neo home robot embedded into a rebuilt version of its marketing site. Each run follows the same loop: research first (architectural drawings, product photos, GitHub skills from skills.sh), write a PRD, then hand it to a fresh agent with a /goal command and let it run for 20 to 54 hours straight, spawning its own research agents, pulling free assets from PolyHaven, and using computer-use screenshots to QA its own geometry. Results range from genuinely photoreal to blobby and unfinished in the rooms it rushed. Estimated total cost if billed via API: about $1,700, fully covered by the $200/month ChatGPT Pro plan.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Pat states the premise: give GPT-6 Astra control of Blender and build three things with zero Blender knowledge of his own.
00:25 – 04:27
02 · Build one: Fallingwater
Pat riffs an unstructured prompt live, telling Astra to recreate Frank Lloyd Wright's Falling Water pixel-for-pixel, research first, and produce a PRD before building.
04:27 – 06:25
03 · Finding a Blender skill
Astra compares the Blender MCP bridge against the raw Python API; Pat points it at skills.sh to find a highly-starred Blender GitHub skill instead of starting from scratch.
06:25 – 09:10
04 · The PRD, then kicking off the build
The research agent delivers a detailed PRD covering measured drawings, a photo inventory, and realism methods; Pat opens a fresh agent and fires it off with a /goal command.
09:10 – 12:05
05 · 21 hours of nonstop building
Astra runs continuously, using computer-use screenshots to QA its own geometry, pulling free assets from PolyHaven, and researching a specific tree species for the landscaping, all without being asked.
12:05 – 12:31
06 · Fallingwater is finished
After a day, five hours, and eleven minutes, Astra stops. Pat inspects the model directly in Blender while his RAM strains under the finished file.
12:31 – 14:27
07 · The flythrough
Astra's rendered walkthrough shows photoreal early rooms with a few rendering glitches, and visibly rushed, sparse later rooms where detail and lighting fall off.
14:27 – 15:46
08 · Where this is actually useful
Pat pitches the obvious commercial angle: real estate, architecture, and interior design firms, plus a cold-outreach idea of underpricing 3D renders to developers found on Zillow.
15:46 – 17:05
09 · Build two: my studio
Pat turns the same process on his own recording studio, using a single shaky 360-degree phone video as the only visual reference.
17:05 – 19:52
10 · The studio prompt
Astra stitches frames from the walkthrough video into a contact sheet, tone-maps the washed-out HDR footage, asks clarifying questions, and writes a studio-specific PRD.
19:52 – 20:43
11 · Watching it correct itself
Mid-build, Astra flags its own errors out loud, a wrong furniture angle, a door in the wrong spot, shelves that need to move with it, and keeps iterating.
20:43 – 22:52
12 · The finished studio
Pat has to ask Astra how to navigate Blender before he can actually view the result. Once he does, the console area is startlingly detailed while some corners stay boxy.
22:52 – 23:12
13 · Asking it where the shelves go
Because the studio is now a live 3D model, Pat can ask Astra to preview new shelves or decor on the back wall before he buys or installs anything.
23:12 – 24:13
14 · Build three: the NEO robot
Pat switches targets to 1X's Neo home robot, aiming to recreate the product page's branding and generate a 3D render of the robot for the site's hero section.
24:13 – 27:50
15 · The NEO prompt
Astra is told to extract 1x.tech's branding, colors, and layout directly from the live URL, gather reference photos and video of the robot, then write a PRD before building.
27:50 – 29:58
16 · The 3D robot
The robot build runs roughly 20 hours and 69 revisions, checking itself against the official reference each pass, before Pat manually tells it to stop once it's close enough.
29:58 – 30:26
17 · Building the website around it
Pat asks for three different creative directions for embedding the 3D robot into the site as visitors scroll, rather than a plain 360-degree spin.
30:26 – 33:15
18 · Scrolling the finished page
Pat picks the strongest of the three scroll concepts, cuts a redundant 'keep scrolling' prompt and a repeated video clip, and Astra rebuilds the full page around the chosen interaction.
33:15 – 35:08
19 · What all three cost
Pat estimates roughly $1,700 in combined token cost across all three builds from the raw input and output counts in the build logs, all absorbed by his $200/month ChatGPT Pro plan with no separate API bill.
35:08 – 35:37
20 · Outro
Pat points to a companion blog post with the Blender skill and full prompts, and invites viewers to try their own builds.
Atomic Insights
Lines worth screenshotting.
GPT-6 Astra ran one Blender build for 24 hours and 16 minutes without stopping or hitting the $200/month plan's usage limit.
The Neo robot render alone took 69 revisions and roughly 20 hours before the agent judged its own work close enough to the reference.
Astra connected to Blender through the official MCP bridge with zero manual configuration required from Pat.
The recommended setup is hybrid: the Blender MCP for live inspection and corrections, saved Python scripts for precise, repeatable geometry.
Astra used its own computer-use screenshots to catch mismatches, like a writing desk and chair that should have replaced a long upholstered seat.
When reference material ran short, the agent pulled free 3D assets from PolyHaven on its own initiative, without being asked or given explicit permission.
It also browsed the web mid-build to look up reference photos of a Canadian hemlock tree so the landscaping would render accurately.
Running research and PRD-writing as a separate first pass, then handing the PRD to a fresh agent with a clean context window, produced far more specific build instructions than a single riffed prompt.
skills.sh works as a ranked directory of pre-packaged GitHub instructions for using specific software, here Blender, that any agent can be handed as a head start.
Estimated total cost across all three builds, if paid through the API instead of a flat subscription, was roughly $1,700 in tokens.
The studio scan was built entirely from a single low-resolution, badly lit 360-degree phone video, with the agent tone-mapping the HDR footage itself before judging room colors.
Later rooms and the tail end of each build consistently look worse than the opening sections, because the agent spends disproportionate time on whatever it tackles first.
Takeaway
Long Unsupervised Agent Runs Need Structure, Not Just Time
AGENT WORKFLOW
The research-then-PRD habit, specific corrective feedback, and an explicit stop condition are what turned days of unsupervised Blender generation into usable output instead of an expensive, unbounded loop.
02Build one: Fallingwater
Riffing an unstructured, repetitive prompt out loud (saying 'realistic' five times) is fine as raw input, as long as a later agent turns it into a specific PRD before building starts.
Naming a real, thoroughly documented reference (a historic house with existing blueprints and photo archives) gives a build agent far more to work with than a vague description.
03Finding a Blender skill
Before letting an agent touch unfamiliar software, have it compare its access paths (a live control bridge vs. a scripting API) and pick whichever produces the better output, not just whichever exists.
A ranked skill directory like skills.sh turns 'has this agent used this software well before' into a searchable question, instead of hoping the agent invents best practices from scratch.
04The PRD, then kicking off the build
Splitting research and building across two separate agent sessions, so the build agent starts with a clean context window and only the finished PRD, keeps instructions sharp instead of buried in the back-and-forth of the research chat.
An agent can connect to specialized software like Blender through an existing MCP bridge with zero manual configuration, when that bridge is already present in the environment.
0521 hours of nonstop building
A single agentic run went 21+ hours without hitting a flat-rate plan's usage limit, which changes how long you're willing to let a build run completely unsupervised.
Left alone, the agent pulled in outside resources on its own initiative, a free 3D asset library and reference photos of a specific tree species, that were never explicitly requested.
06Fallingwater is finished
A 24+ hour unsupervised build still finishes unevenly: whatever the agent tackles first gets disproportionate polish.
07The flythrough
The rooms an agent builds last in a long run are reliably its weakest, less detail, worse lighting, more visible artifacts, because the token and time budget runs out before quality does.
08Where this is actually useful
Turning a capability demo into a business idea can be as simple as finding live listings for sale, contacting the developer or broker directly, and offering a 3D render at a steep introductory discount to build a portfolio.
Architecture, real estate, interior design, and construction are the obvious first buyers for AI-generated 3D walkthroughs because the output maps directly onto a deliverable they already sell.
10The studio prompt
A single messy, badly lit walkthrough video is enough raw material for an agent to reconstruct a real space, as long as you also hand it explicit measurements.
An agent can correct for bad source footage itself, tone-mapping washed-out HDR video before judging colors, rather than needing to be told the footage is unreliable.
11Watching it correct itself
Early passes come out cartoonish and blobby by default; specific written feedback about proportions and object placement pushes it toward realism faster than leaving it alone longer.
12The finished studio
When an agent doesn't recognize a specific product in a scene, telling it to go look up product photos online works better than trying to describe the object in words yourself.
Whatever the agent iterates on longest ends up disproportionately detailed, here that was the desk console, while other corners of the same room stayed boxy and approximate.
13Asking it where the shelves go
A completed 3D scene isn't a one-time output: because it's an editable model, you can keep querying it, asking how a new object would look before you buy or install anything in the real room.
14Build three: the NEO robot
Feeding an agent a live reference URL, not just a description, lets it extract exact branding, colors, and layout instead of approximating them from memory.
15The NEO prompt
The same research-first pattern holds even for a much smaller build: gather reference images and video, then write the PRD, before ever touching the modeling software.
16The 3D robot
An agent can keep iterating, 69 revisions over roughly 20 hours here, well past the point of obviously diminishing returns, so a plain 'that's close enough, stop' instruction is still necessary to end the loop.
A single-object generation can take just as long as an entire building if the agent is never told what 'good enough' looks like.
17Building the website around it
Asking for three distinct creative directions, instead of describing one exact interaction up front, produced a noticeably better final result here.
18Scrolling the finished page
Being explicit about what to cut, redundant 'keep scrolling' text, a repeated video clip, mattered as much as describing what to build.
19What all three cost
Without a built-in usage dashboard in the coding harness, estimating true cost means manually approximating it from the input and output token counts sitting in the build logs.
The same workload that would run roughly $1,700 in raw API credits was fully absorbed inside a flat $200/month subscription, which is the actual argument for the higher-tier plan over pay-as-you-go API access.
Glossary
Terms worth knowing.
PRD
Product Requirements Document. Here, a written spec a research agent produces before a build agent starts working, listing exact measurements, materials, and reference sources.
MCP (Model Context Protocol)
A bridge that lets an AI agent send commands directly into a running application like Blender, instead of only writing code that runs separately.
Extra high effort
A reasoning-effort setting for the coding agent that trades speed for thoroughness. The video deliberately avoids the top 'max' setting because it can behave erratically.
skills.sh
A directory site for ranking and finding GitHub repositories packaged as reusable instruction sets, called skills, that an AI agent can load before starting a task.
PolyHaven
A free library of 3D assets, including textures, models, and lighting setups, that the agent pulled from on its own to speed up the Falling Water build.
/goal
A slash command in the coding CLI that hands an agent a task and tells it to keep working in a loop, queuing its own steps, until the stated goal is met.
“Much more worth it when paying $200 versus $1,700 in API credits.”
the exact cost-justification punchline→ newsletter pull-quote↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphor
So today I'm giving the best coding model in the world, GBT6 Astra, the power to use Blender, the free 3D modeling software. And in this video, we're going to be building three things with it. A fly through of Frank Lloyd Wright's Falling Water, the room I'm sitting in right now, and a 3D model of the Neo Humanoid Robot and actually embed it into a website.
You don't need to know how to code or anything about Blender. I certainly don't. This setup takes five minutes with the right prompts and directions I'll be giving you in this video.
You'll be making these 3D models in no time yourself. So without further ado, let's dive right into the first build, which is Frank. Lloyd Wright's Falling Water, which if you're not familiar is a home in Pennsylvania that Frank Lloyd Wright designed in the 30s.
It's beautiful. It's built right on top of a waterfall. And I really wanted to see GPT -6 Astro bring into life with Blender.
So that's what we're going to do for this. And what I'm going to do as well, I'm not going to paste in a pre -written prompt. I'm just going to try to riff here just so you can see my thinking and see how easy this is.
Again, I have no knowledge of Blender. All I have is the app downloaded, which you'll need to do. You can just go blender .org, download.
download it and then we're just going to tell astra to access blender so let's do that now we'll go over to astra and i'm just going to start riffing so i want you to recreate frank lloyd writes falling water basically pixel for pixel in blender i don't know how you access the ui i'm assuming it's through the mcp server maybe it's through the api but i need you to access blender control it completely and build out this falling water home to an absolute t i'm talking as realistic as humanly possible it should look like from afar from first glance like it is real.
That's the level of detail that I want you to match. I want the lighting to be beautiful, every single minute detail. And what I want in the end is an ability to look around the exterior and then actually fly through the different rooms.
So I want each one of these rooms mocked up properly, built properly with Blender. I think there are quite a few blueprints. There are a ton of photos, of course.
I think even the Library of Congress has a lot of this information available. So do a ton of research first, gather all of that, make sure that all of that research will feed into the build. find everything you need to before you start building and then i need you to do more research into how to actually make this as realistic as possible like do not stop in your researching you until you know exactly how to make this freaking beautiful in blender i don't want any blobs anywhere and then of course i need you to actually qa your work so do all of that research you can spawn different agents put all that together first then i want you to actually put together a prompt .md you can add it to my folder here and we can treat this as a prd and then what we're going to do is once you're done with all of that, then I'm just gonna do a slash goal.
And then I'm just gonna say, read the prompt .md and execute on all of this. So let's kick this off. We're gonna go extra high effort.
I don't wanna go max because that could get a little crazy. We're gonna go extra high. We're gonna kick this off.
Let me explain a little bit of what I'm doing here. First. said realistic about five different times i don't know if that's really necessary but i'm just meta prompting so what i want is to have a just cleaned up prompt because this is kind of a riff so i'm having this agent just kind of function as the research agent put all this stuff together spawn other agents and then i'll probably open up a new chat and just have a separate agent with new context just execute on this whole prompt md with slash goal so it's going to go into a loop it's going to make sure that it's queuing all its work and it just won't stop until we get this as realistic as possible so we'll let it do that research i'll show you the prompt that it comes up with because I will need to review that.
And then I am curious too how it's going to access Blender. So I'll talk more about how it actually does that as well and when it gives me an answer. And then we'll kick it off with a new agent and actually start building.
So, okay, we already have a question. Which version should the reconstruction represent? I'll use the restored house with its documented furnishing in the late summer surroundings unless for historical reconstruction.
Okay, yeah, late summer setting, nice. We'll probably have more information on that anyway. And then it looks like, so just answering this Blender question, Blender 5 .1 .2 is installed.
I can control it directly through the Python API. there isn't a Blender MCP server configured. So I would need to configure the Blender MCP server.
I'm gonna ask if the Blender MCP gives any benefit, if there's any more features that the MCP will give over the Python API. So I'm just gonna ask Codex that now. Okay, well, we already have Blender popped up.
Let me just go back to Azure here and ask this. So question for you on the Blender MCP versus the Python API. Can you just do research and just make sure that if you do use the Python API, is there anything that the Blender MCP server does better?
If so, like if it gives, you any more features if it makes this quicker, more efficient, makes your output any better, then use the Blender MCP and tell me how I can configure this. Otherwise, you can just use the Python API.
The other thing too I would recommend is just there are some open source libraries. There's some GitHub repositories that you can access that will show you how to use Blender better. You can check out a site like skills .sh and find some of these highly rated, highly favorited, highly starred GitHub repos in skills .sh.
When you spawn those research agents, have them look that as well so i'm just going to steer it and then what i'm telling it to so skills .sh if you're not familiar this is usually my go -to for any skills if i'm using any program i just start with skills .sh and see if something exists so when i search blender i've done this before i know there's quite a few blender skills you can see installs are really high security audit is safe github stars only 79 github stars i think there's some better yeah here we go 251 000 github stars this blender motion state inspection so i can just give this to an agent but i'm gonna have codex just do all this research if it comes back confused or if it misses this in any way then i'll just give it this but i want to just give the research agent freedom to just go and find the best one but On first glance, I think this is the most efficient.
And then there's also like a Blender MCP, but you can see the GitHub stars are kind of low. So I'm just going to give that agent the freedom to make its own decision and we'll see what it comes back with there. As far as supplying skills, this isn't totally required.
I do like to just have something for an agent to start with because people have spent some time organizing all of this and how to call Blender, you know, the most efficient way to actually use Blender for an agent. So let's go back here and then just let it keep running. Okay, according to Astra to answer the Blender MCP question, it looks like it's recommending a hybrid approach.
So the Blender MCP for live inspection and corrections with saved Python scripts for precise repeatable construction and rendering. By the way, I'll put together a skill. I'll make it available in the link below just based on my learnings as I go through this process.
So you can just take this, hand it over to an agent and that way you're not starting completely from scratch. So I'll have that available. I'll also have all the prompts that I give each one of these builds.
That'll be available on a blog post as well for you to just grab. You can copy this if you'd like, if you want to do something very similar. And we're just at the research step right now.
It looks like it's putting together a reference package with 15 measured drawing sheets 49 interior exterior photographs so this is what it's going to use to actually put together the model and blender all right astra has been humming for about 40 minutes it is slow but it is efficient looks like the prd has been created it's reviewing the prd first and so once that prd is written we'll take a look at it okay so we have the prd put together by our first astra research agent so let's take a look at this what i'm looking out for specifically is making sure that the research agent has included all the terminology we need not that i actually even know what this terminology is but if i see a lot of keywords a lot of things for an agent to know exactly what to do that's a good sign because it's much easier for this next build agent to follow specific direction as opposed to just saying something like make this realistic which is another reason why i started with research first and i didn't have it start building from my initial prompt because when i say make this realistic it's just not specific enough so that's what we have laid out here and let's just take a look at this prd intended results build franklin writes falling water and blender including the main house guest houses and staff
Service spaces connecting circulation, terraces, pools, waterfall, rock formations. Yeah, see, this is the type of stuff that I wanted to define initially with a separate agent. So surrounding woodland visible during exploration.
The visual ambition is photographic credibility at first glance from a distance with carefully observed geometry and materials that hold up at human viewing distances. It goes on, user confirmed setting, restored house with documented furnishings in late lush summer. Read and use this research.
So here's the research. So we have different files now within this folder. the research done architectural evidence every space inventory and dimension anchors wow okay so we're getting like into even furnishings and materials and everything that is in these rooms blender realism methods okay yep so i did a bunch of research on how to get this realism in blender no surprise astra put together a really good prd now what we're going to do is we're just going to kick off this build so i'm going to go new chat i'm going to go slash goal i'm going to say build out rank load rights falling water pixel for pixel follow the prompt .md exactly dude not stop qa all your work once complete do not come back to me until perfect like it really doesn't matter what you say here you could i could just say follow the prompt .md and it'll know i'm just going to mention too like this is obvious but i'm just going to say all the research docs are in this project folder as well and we'll kick this off and probably come back in 11 hours okay cool so we've got the build kicked off it looks like blender's responding through the official mcp bridge to be clear i didn't set up anything there's no mcp configuration that i had to do my
so it looked like Astra was able to just go into Blender here and access via the MCP.
All right, well, it is the next day. I knew this thing was going to be running a long time. I did not know it was going to be running this long.
It has been 21 hours of nonstop work. I mean, you want to talk about Astra as a workhorse. This is just insane.
I'll check in on usage at the end. I'm on the 20x $200 month plan, but it doesn't seem like I'm hitting any sort of usage limits or anything like that. And it's been running for 21 hours straight.
If we check in on progress, I mean, it is being incredibly efficient. I'm just going to spoil some stuff. Look at this attention to detail.
I mean, come on. Look at the look at the lighting. the chair there so we went a little over the top i'm just absolutely roasting tokens i'm realizing that there are a lot of rooms in this place and there's a lot of detail that needs to be added we'll see how accurate it is to the real thing but i mean on first glance this is exactly what i was looking for and astra is absolutely crushing it and if we just look at what's been running too just to show you hide this just move this over it's just using computer use to actually check its work here so this is everything keep going keep going there we go this is the prompt that started it and it's just been working So it's doing things like checking the furniture.
I also confirmed another furniture mismatch in the guest room. The southwest corner has a writing desk and a chair, while the current model has a long upholstered seat. I mean, that is the attention to detail that this model has.
It's just, I've never seen anything like this before. It really is incredible at this. And then it looks like it's doing something.
I don't even know what it's doing here, but it's pulling up. i guess polyhaven so polyhaven must be some sort of free library where you can download some of these assets so it is cheating a little bit it's not it's not generating these from scratch but the fact that it even did this it didn't ask me for permission to do this or anything it just knew it had to get done to make this look realistic and so it went looking for that itself so we're gonna let this keep running for a little bit longer but i really need to get this done and move on to the next example i mean look at again look at this this is what it's doing with browser here it's looking at a canadian hemlock tree at images of it to make sure that it renders it properly in blender so it is accurate to the trees i don't know where the trees are i guess outside i mean i'm just blown away by what this model can do again this is just so cool now it's looking for more rooms and it's doing a bunch of research like this this guy cody james photo looks like he's taken some photos in the falling water house like this photo so it's taking this photo i didn't tell it to do this i don't think this was in our research doc either it just felt like it was still incomplete based on our research so it's doing even more research
these separate photos extracting these and then turning them into renders in blender i i just cannot get over how good ashra is at this all right so we've officially reached the one day mark with this falling water generation 24 hours and 16 minutes of generating we have let's see a bunch of staff rooms generated we've got some stone and wood generations and some rooms i think it was working on the exterior at one point i really want to make sure it nails the exterior but it seems to be gravitating towards getting these rooms right because there's so many rooms I don't know how many rooms are in this place, but it seems like there are a ton because it's just constantly working.
It's going room by room and just tackling each one of these. Then it's QAing and it's noticing that it's not exactly accurate. Then it's just going around in a loop.
And I was going to tell it to just wrap up, but I kind of want to see just how long it will keep going for and see how close we can actually get when we do this fly through. Okay, one day, five hours, 11 minutes, and three seconds later, Falling Water is complete. So here it is in Blender.
It looks amazing. There's these little artifacts here. It just has to do with the way it's rendering.
my RAM is just getting absolutely hammered right now. So it's very glitchy. Even when I click play working tour, it just isn't rendering properly.
The actual camera movement is very weird, but the actual generations, I mean, look at these rooms. It looks amazing. So what I had Astrid do was just put together a video.
So it rendered this full Falling Water tour. And if we go here, it's looking really good. So here it is a full screen and just play this.
There's a few little things, like there's still some artifacts, just these grainy artifacts. There's like a little bit of flashing going on. The water looks a little weird, not exactly realistic, but I mean, the actual like 3D work just looks really good.
The lighting, the rooms, I don't know how accurate this is, the actual rooms within falling water, but judging by the amount of time it took for Astro to generate this and the amount of loops, I think he got it pretty close. Some weird things here and there with just these bowls hanging.
Granted too, I mean, this is just a really difficult generation overall. Like if you're doing this yourself with some of the use cases, which I'll talk about in a second, you should be able to do this much quicker. So this, and then this is where it just kind of gave up because I told it, we just got to get this thing done.
So the rooms that are kind of sparse right now, this next room doesn't look bad. Kind of a weird camera angle. I think just zoom out just in capturing the actual render.
And then this, it starts to lose the detail towards the end. And then we just have like these highlights blown out here at the end. Same thing with this other room, just not that much detail.
And then not a whole lot of detail here. But these beginning generations, I mean, they look fantastic. And this is the whole point just to show.
what it can do with 3D work, but you can see how long it takes for something like this. So that is more or less a pixel for pixel recreation of Frank Lloyd Wright's Falling Water with GPT -6 Astra. I mean, you can see just how crazy this is, how long it runs, good it is at 3D with Blender, and just how far these models have progressed.
I mean, even a coding model from a couple months ago, you would not have gotten close to these types of generations. But however cool this may be, you're probably thinking, okay, but how is this actually useful? So I did give some thought to how you could actually use this.
your day -to -day so here's a couple use cases that just came to mind with these types of generations i mean of course the obvious ones are real estate architecture and construction if you're in these industries or in these adjacent industries this would be incredibly useful for clients for your firm interior design too these are the more obvious ones developers stuff like that i'll talk more about industries outside of this but just its ability to take a building take photos and completely recreate it i mean this could be a killer upsell if you're in one of these industries and then of course i actually selling to them so what you could do i'm just i'm just riffing out loud i know this is not what this video is about but you could go to zillow you could find homes that are for sale you could contact the developers and say hey i have a 3d render that i did i think it would be really useful for your clients could i sell this to you for i don't know start super cheap a few hundred bucks people probably pay thousands of dollars for this but just undercut the prices at first because you're just running with astra and you can just do this at scale i mean it's really easy to contact brokers and developers their emails are right there and you can just start generating these like you just
click into this and you just say hey astra look at this virtual staging take these images and recreate these in blender so that's just one idea i wanted to not just show off how cool this is but also just give some thoughts how this can be useful now of course having said that this still may not be totally relevant to you so let's talk about something even more useful all of us of course have a space that we're either living in or working out of or both so what if we could mock up ours let's do this next and what i want to do is i have this new studio space and i'm still trying to figure out the lighting camera placement decorating it a little bit even where to place some of the decorations especially for the camera in the background so i have all these measurements in a video that i took which is here It's a little blurry.
It's just this 360 view. Don't look too hard. The studio is a complete mess.
We have this whole studio here. And so I shot it from all these different types of angles so it gets everything. I'm going to give Astra the measurements.
And then I'm going to see how close we can get to a real render in Blender of my studio space. Then I can ask Astra questions. What if I place this here?
And I'll be able to tell exactly what it looks like before I actually put up floating shelves and stuff like that. I'm going to go over to Astra on extra high again. And I'm just going to riff a little bit.
so the goal with this is for you to mock up a pixel or pixel recreation with precise measurements of my studio space i have a video of the space in this folder and it's kind of all over the place because i couldn't get a proper 360 view of everything so you're gonna have to kind of stitch this together take pieces from different angles different parts of the video and put it together into maybe a contact sheet at first just to get a general sense of the space as a whole i think i got all the different angles and stuff but just check that and then i will give you the measurements of the space as well you understand that and what i want you to do is mock up this entire space with blender i need it to be really really realistic too no blobs or anything like that the lighting should be correct all the different pieces of furniture the carpet everything should be a pixel for pixel recreation at first glance i should think that this is a photo of my studio that's how good and realistic i want this render this generation in blender to be so go ahead and just do a research pass first gather all the information put together a prd let me know if you have any questions
and then we'll get going on actually rebuilding this studio. Cool, so we'll go ahead and just send that off and see what this research path will come up with. Okay, so it inspected the video.
Video is HDR, so I'm making properly tone -mapped reference frames to avoid judging the room. From Washed Out Colors. Good call.
Yeah, because that video did not look great. Looks like we've got a couple questions for me. What measurements do you already have?
I forgot to give my measurements. First, finish recreation. Preserve the room exactly as filmed, including boxes, cables, cloths, water jugs, and the current daylight.
Yeah, remove temporary clutter. That's a good call too. Sweet.
So we have a contact sheet mocked up. Check this out. Did a pretty good job of this.
just took screenshots from this video and then we have this whole prd intended results created an editable dimensioned blender recreation of pat's actual studio that reads as a photograph at first glance match the room envelope furniture silhouettes materials rug patterns equipment an existing mix of daylight and artificial light wow very good okay then we have the specs we have all the measurements room and object scope what's required to make this look realistic so we're going to do the same thing we did with falling water and i'm just going to take this and i'm just going to create a new chat an agent with fresh context and get it started so new chat we're going astra extra high i'm just going to go slash goal read the prd in the project folder and recreate studio in blender pixel for pixel qa all your work be incredibly thorough don't stop until goal is reached again unnecessary i could just go slash goal say follow the prd and it would do the same thing but i just like to emphasize this all right astro is humming on these generations here is its first pass pretty good i mean this is like blobby and it's even giving itself feedback right now it's realizing that texture lighting furniture proportions several concrete layers the lounge chair needs a different angle the door is rough roughly half a meter further along the wall utility shelves need to move with it i'm even going to give it feedback just to make sure i think it knows
this but it needs to look much better this looks very cartoonish right now i just gave him more feedback too and i said Because it was asking me about what keyboard, mouse, stream deck models do you use? I said, I don't know.
Just look at the video. Try to find which one this is. Do this for all the objects in the space.
Just look closely at it. Try to find out what the actual product is. And I don't know if you go and look up images of relevant products.
So you can mock it up that way. Or you can just take the video. Just try to match what's in the video.
But it might actually need to look at like product photos for these objects. All right. At the one hour mark with the studio build.
Looks like the product shapes are improving. We will see about that. And the next checks are about.
they sit together here's a couple preview images what this is looking like yeah looking much better still kind of like cartoonish and blobby we're gonna fix that but i mean this is pretty pretty darn good for a first pass okay so we are at 3 hours and 37 minutes with this studio mock -up looks like it's making some good progress it's finding a lot of errors it's being very thorough as astra always is like you can see how detailed it gets again just like the frankfurt right generation moving the console about 11 centimeters forward improves all three lounge views while keeping it grounded it's like kind of attention to detail almost too much attention to detail because all i really care about like at least for this is for it to look realistic it's getting more into like the sizes and stuff but sure why not i'm just gonna let it run all right in about three hours it looks like it completed three hours and 34 minutes we have a render of the studio i'm just asking astra how i actually navigate blender because all i see is this which this looks pretty good but i don't know how to look around in 360 view because i know nothing about blender i took a minute to actually figure out how to navigate blender
but check it out, here is studio. This is actually really impressive. So, okay, as we zoom in, we've got the monitors here, the lights, cameras, we've got the teleprompter.
Some details better than others. I mean, granted, it took this from a video, a low -res video. It's got the coffee maker in here.
It's got disinfectant wipes. It's even got my cafe caribe coffee. the actual background.
And then here, especially with the console, this is where it's really impressive. So check this out. So this is the console here.
And then here is the console. Actually, like that is crazy attention to detail here. And I think that's where it spent the most time too.
I did kind of have it, had it rushed the rest of this. We can see that detail here. And if it worked for, you know, another few hours, it would get the detail and everything else as well.
Like some of this is boxy, kind of inaccurate, but. I mean, some things, it even got like what was on my monitor at the time. It even mocked that up somehow.
Grid's looking good. I have this second camera here, teleprompter, another monitor, light here, these bins. I mean, this is really, really impressive.
Like the detail on this bamboo, these lights, Google Home, this Polaroid. I mean, it has everything. It even has this like 3D camera that I have down here.
Just wild. Lens, a book. it has the book title ugly ugly on advertising the mesh of the chair geez that look at this detail here so that is the generation of the studio and like i said if i really wanted it to i could have it keep going with details and the best way to do this if you're doing this yourself just take pictures of the objects or have it look for photos of the objects online and it'll mock it up that way because some of this stuff like it was just pulling from my blurry grainy video so it couldn't really get all of the detail but when it had that like the console like the chair see just how realistic it is just incredible and this is all in blender this is a complete 3d model now so what i could do as well is i could say mock up how something would look you know on this back wall because i want to hang up some lights and floating shelves and i could even say like okay here's my camera i could give the height of my camera and i could say okay now tell me what it looks like on camera insane insane astra is a beast so that is building a space you're in in 3d but of course you might do this one time then never again so for our last build i wanted to do something that will be relevant
to anyone with any kind of digital presence. So say you've got a website and you're looking to spice it up. Arguably, one of the best ways to do that is some kind of 3D generation.
This is becoming super popular nowadays and it always makes a site look premium and stand out. And this is absolutely where website design is moving. So let me show you how to do this in Blender and then embed that generation in a website.
So for this idea, we are going to build this for Neo, the robot. I think it's called Neo Robot. I forget what it's called.
One X Neo Home Robot. so here it is pretty cool inside already but what we want to do is take this robot make a 3d render of it give it some kind of rotation make it super interactive engaging on their website and maybe we'll pull some of these elements too even like these videos and stuff i like these cards i like the thought of trying to recreate this but keep it on brand with them and just add our own kind of 3d elements to this i'm surprised they don't do 3d here actually let me see like this would be a perfect example so that's what we're gonna do and just like we did with these last prompts i'm just gonna go over to astra and i'm just gonna riffing so i want to recreate neo it's a robotics company 1x .tech neo here is the url extract their branding as much as you can even the way they're doing like this like studio kind of background with this header image their nav bar like i really want you to match this even pull their logo these rounded corners pull this video too this little glass morphism ctas icons let's try to just copy this as much as we can but our goal with this is to generate
a 3D model of their robot here. And that's gonna be the hero image. That's what's gonna show up first.
And as we scroll, make this interactive somehow. Make this incredibly engaging where as I scroll, maybe it's the robot turning around or maybe we embed some things like what the robot's actually doing. But basically I want a 3D render of this robot first.
Use Blender. the MCP and the API and then go ahead and just generate this. And like I said, I will leave it up to you where what you do with 3D.
I really like this image at the end. That's kind of what I picture for this, where it's like the studio background. There's this nice like opacity fade at the bottom.
But what I need, of course, like I said, once again, full 3D. So complete rotation, everything. And like I said, I'll leave it up to you on how the website actually interacts as you scroll.
But I need it to feel super premium, super on brand. So you're going to need to extract multiple images of the robot to get. this and probably call out some product features too that's my long riff and then the last thing i'm going to say is do a bunch of research first find out the best way to do this put together a prd based on all of your findings in addition to the website the neo website that's been extracted then i'm going to hand this over to another agent so i want you to put together the prd and then i'm going to do a slash goal command with another agent we're going to actually build this out so again meta prompting here we're just going to kick this off with this agent here that's going to function as a research agent and then we'll hand it over to another astra to actually do the building All right, so we've got the PRD, this reference pack here.
It basically just put together a website, which is fine. This looks great. It's got the rounded corners of these images.
It extracted all the different reference photos of the actual humanoid robot. Got closeups of the legs, the shoes. It's got these extracted videos.
It pulled the iconography from the website as well. All of these images. So everything is ready to go.
Let's just take a look at the PRD as well. Assuming it's fine, just quickly glance at this. Yep, objective and delivery order.
premium update to the existing 1x neo page build the robot first show the 3d results and we have a bunch of findings and we're using our blender skill to generate these just get as realistic as possible i mean it's got everything in here no surprise astra put together a really good prd so we are ready to kick this off with a build agent and then for this actual build agent we're going to go slash goal and i'm just going to do this just paste in this prompt basically what i'm saying here is follow the prompt .md slash prd in this folder create the neo robot 3d generation again this isn't necessary all this is already laid out i just like to be super specific with all this there were some things that i thought of like we may not have all the angles from our research just as i was looking at those photos again so i just said fill in those blanks and then i just say i mean this is already in the prd but check in with me once the robot's been generated and then i just say use the blender mcp and the blender api etc etc so we'll just kick off this build here and see how long this one takes uh okay you gotta be kidding me here we go again this thing's been running for 20 hours i just let
it run overnight. I had some work to do this morning, so it's been humming. It's impressive, but it's also a little annoying because, I mean, a 3D generation like this shouldn't take that long a time.
It's not like it was doing Frank Lloyd Wright's entire falling water home room by room. Like, that's justifiable. This seems a little crazy to be running for 20 hours and 22 minutes.
However, the actual generation, this is the generation. This is not from Neo or anything. This is actually what it created, which is pretty darn impressive.
I mean, you can see the mesh here. all the texture shoes the face everything looks freaking good and so you can see it went source and then review 68 study 69 so it's at the 69th iteration of this here's the full history so this is all yesterday yesterday today 2 22 a .m today 33 a .m 4 a .m 4 43 a .m revision 58 and it just keeps going and going and going and going and going until we are finally at this afternoon we're at the 69th revision of this.
I think this is good. I mean, look at this. Like even just the difference.
This is review 68 and then study 69. It just got like even more detail. It's just amazing.
So the source, it's matching the color. So you can see right now it has this kind of cream color. This is white.
This is like getting more towards the cream color, but it's fine. Like we're close enough here. So I'm just going to stop it.
Okay, great job. You can stop now. It is close enough.
You can stop now as long as this is a full 3D. model can i can i just play around with this let's see open the interactive robot let's see what we got here okay i guess oh it's right here all right cool let's check it out official one x reference our version pretty cool wow it even got the thing on the back assuming it's some sort of charger or something you can zoom in on it pretty impressive like really really impressive use the complete model then we go front left side hand wow So it has like all the references of the hands.
That's really cool. Shoes, face. That is really cool.
And it's very close. It did a really good job with this. Like even look at all the texture and the lighting here.
It looks so realistic. Like even in here, it has the reflection of like a studio light. Just insane.
So this thing is ready to go. Now what I'm going to say is, all right, the 3D render looks perfect. Well done.
Now the question for you is how we embed this in the website. I want it to look premium, professional. I want it to be engaging.
I don't think it's just a 360. view of the robot. I think it's something else, like almost, I don't know, see if you can do something creative with the scroll.
I want you to basically recreate their website. We have a lot of these assets in our research already. Recreate their website and just embed the robot somewhere, probably in the header.
And I just want you to play around with, give me three different options on how we can do this. Remember, I don't want it to just be an ability to turn it around. Think about this robot as a whole.
Think about what it does. What can we do to show off its features in a way that's engaging using this 3D render somehow.
Remember, three options, recreate the website as well, then I'll give you feedback there, then we'll finalize the website. All right, so we've got some 3D interaction options. First, this one, Meet Neo.
Let's check it out. Okay. Oh, nice, nice, nice.
I like it. I like it. Yeah, this is nice.
good call astra uh nice we've got a video in here that's a nice touch it's just embedding that video keeping on brand with these rounded corners the text font all of that this like next feature that's kind of lame get rid of that okay and then it's just the same video but okay i think oh yeah i see what it's doing so it's just showing it's just showing the first beginning generations uh that one's pretty good let's see what the other ones are okay day together morning afternoon evening connect the robot interesting like be getting a little too creative here but all right so we're scrolling all right same kind of thing so yeah this is the okay that's not bad though kind of a smart idea like it really did think i mean this is such a creative model like at least it's thinking outside the box and it's really considering the product features like you would you know when you're building a website like you know the fact that it can do all these different tasks throughout the day so it's like you're kind of scrolling throughout the day uh but i just don't love the just a little plane but i see what it's doing so the idea i think it nailed it with just this you scroll to uh the difference like
the lending the hand or whatever it's whatever it's doing for that first one i think that's fine but let's just take a look at this last one it's the same kind of thing you scroll in expression yeah okay so so the first one's fine all right perfect first one is approved build out the rest of the site just get rid of the the like next feature or pause motion like it's implied that you keep scrolling you don't need to say that again and then make sure you don't repeat the video so it's you know now imagine your day then transform your home that's the same video i'm assuming you'll have a new video and then scroll to meet neo you can just get rid of that like it again it's implied with the scrolls all right updated website is complete for neo robot let's check it out here we go okay the cta is a little weird that's annoying tuner deposit i don't know why it's should match the cta there but okay let's just keep scrolling made blend hands more than a helping hand cool nice got a video there okay and then we're just kind of like on brand
the rest of the website it's amazing that it just extracted this like this just from localhost this is not their actual website in any way uh it just nailed this formatting and everything automate your chores okay wow it even has this like opacity fade that's pretty impressive but yeah so we have the 3d that's like that's exactly what i wanted just like a nice 3d generation that's embedded in the website like that pretty darn cool definitely makes it definitely makes it more interactive more engaging than you know just a static header image ah and there we go so that is going to do it no surprise this was a long one took about two and a half days of 3d generations across the board with these three builds but i mean you can see how impressive astra is at 3d with blender and just checking it to work too this is something that previous models i mean i hate to keep calling on fable 5 .1 but fable just would it just would have stopped it just would have stopped a couple hours in maybe it might have been faster but it would have taken many rounds of revisions here and that's what astra is really exceptional at so as far as costs go let me just mention these now so here is the blog post that's included in the link below that details all of this you can play through
of these generations i have the blender skill in here the actual prompts for all of these let's just scroll down and get to costs because yeah they're pretty massive as you might expect first up falling water took a day five hours 11 minutes 10 million input tokens we've got some cash reads in there and then we have 2 million output cost is about 674 dollars if i was using the api remember that i'm on the 20x the 200 month plan chat gpt pro so this is all included in usage however i did hit my usage limits at about halfway through the second build and I had a reset banked, so I just used that reset to keep these builds going.
But between Falling Water and then half of the studio, which you can see, I mean, total build time, nonstop build time with that is a day and seven -ish hours. That's finally when I hit my usage limits. And I was starting basically with 100 % usage when I started with Falling Water.
So you can see, I mean, it can run for, at least with this build, it ran for literally over 24 hours before I hit any kind of usage. And then in the studio, I don't know why this cost is so similar. Like I've mentioned in previous videos, videos isn't exactly easy to get an accurate API cost we're just looking at input tokens and output tokens from the build like in the actual logs to approximate a cost there's not an easy way to do this in the codex harness unless somebody knows of one let me know in the comments if there is when I go slash usage it just goes and it just shows my you know regular settings so there's no way for me to actually tell we have the Neo product page that ran for 21 hours and so that costs this is a little bit lower again these are all approximate this is all just estimating based on the input tokens and output tokens the actual build logs it looks like about 440 dollars so total 1 .7k that is a lot that is a that is a lot of money you know that's not something you'd be using the api to do but again it was just covered in my 200 pro subscription so much more worth it when paying 200 versus 1700 in api credits anyway that is going to do it i appreciate you sticking around to the end i hope this video really showed off the potential of gpt6 astro with blender and a reminder too i had no prior blender experience i'd probably only open the app a couple times maybe and you
see the outputs it was giving me so i'm confident you'll be able to do the same if you have similar experience which is no experience and hopefully that blender skill which again is in the blog post link below will help you just kick these off but yeah feel free to do some generations i'm really curious what you come up with send them over just let me know in the comments and i'll see you in the next one
The Hook
The bait, then the rug-pull.
Pat Simmons hands GPT-6 Astra full control of Blender, a tool he admits he's barely opened himself, and lets it run completely unsupervised for over 54 hours combined across three separate builds: a historic house, his own studio, and a humanoid robot's product page.
Frameworks
Named ideas worth stealing.
06:25model
Research-PRD-Build loop
Riff the initial ask to one agent, even messily
Have it research (architecture docs, product photos, GitHub skills) and write a prompt.md / PRD
Open a fresh agent with clean context
Run /goal against the PRD and let it loop until QA passes
The three-step meta-prompting pattern Pat repeats for all three builds to get specific, well-researched instructions instead of a single vague prompt.
Steal forany long-running agentic build task, not just Blender
CTA Breakdown
How they asked for the click.
VERBAL ASK
35:17link
“the blog post included in the link below has the Blender skill, the actual prompts for all of these builds, and the cost breakdown”
Soft, non-salesy CTA to a companion blog post and newsletter rather than a hard product pitch, closed out by inviting viewers to try their own builds and share results in the comments.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
One-shot, no revisions, deployed live: a retro iPod app, a scrubbable D-Day reconstruction, and a Nuketown FPS clone, built back to back by two frontier coding models.
Ten identical builds, five models, blind-ranked before the reveal — a real-world stress test of Moonshot AI's new open-source model against GPT-5.6 Sol, Opus 4.8, GLM 5.2, and its own predecessor.
A creator builds daemon, a single Mac app that runs every AI coding agent he owns, by giving Claude Fable 5 four rounds of blunt feedback instead of writing a line of the UI himself.
Fifteen parallel builds, three models, and a blind reveal system, testing whether Fable 5.1's benchmark gains actually show up in real creative and coding work.
A creator walks through cloning a $2 billion venture-backed dictation app in one Claude Code session, then benchmarks his free local build against the real thing.
00:27
00:55
01:33
02:00
02:26
03:20
03:50
04:18
04:40
05:07
06:00
06:27
06:54
07:20
07:47
08:14
08:54
09:31
09:57
10:18
10:57
11:25
13:11
13:38
13:48
14:26
15:24
16:12
16:31
17:08
17:35
18:28
19:05
19:09
20:09
20:47
21:28
21:59
22:24
23:13
23:41
24:30
25:03
25:25
26:15
26:16
26:56
28:16
28:42
29:09
29:36
30:52
31:23
31:39
32:04
32:33
33:25
34:03
34:42
35:09
26:22
36:06
Video of the Day19:21
13:44
29:59
15:29