GPT-6 Astra reportedly rebuilds a studio in Blender from 5 photos and 3 panoramas in 11 minutes

• Five photos and three panoramas became a Blender reconstruction of a real studio in an initial 11-minute pass, according to LinusEkenstam's original post.
• The reconstruction relied on follow-up questions and supplied dimensions, as LinusEkenstam's reply on measurements describes, rather than image references alone.
• The deliverable reportedly took another eight minutes to turn into a web app, consuming roughly 3 to 4% of monthly quota in LinusEkenstam's runtime estimate.
• Accuracy was uneven: LinusEkenstam's accuracy report identifies errors in the stairs and mezzanine depth, while the kitchen and mezzanine were broadly on point.
• A related Astra-to-Blender setup puts the split at roughly 90% MCP work and 10% computer use when an MCP server is installed, per LinusEkenstam's control-path reply.

OpenAI positions GPT-6 Astra as a computer-use model, but its launch page does not specify the Blender plumbing used in this studio run. An OpenAI Developer Community discussion sketches four possibilities: local app control, MCP/API calls, Python bpy scripts, or direct mesh generation. Gigazine's Blender-demo round-up captures the wider wave of short-prompt Blender experiments.

LinusEkenstam says Astra recreated his studio from five photos and three panoramas, then he shared the Blender files. The published studio viewer also exposes the scene online and offers the model for download.

LinusEkenstam also welcomed an in-person visit from someone nearby in an invitation. In a separate exchange with Alex Reibman and Romain Huet, he said he would take a dev-day ticket in a reply.

The input was a dialogue, not a one-shot image dump. Astra asked follow-up questions, and LinusEkenstam supplied measurements in response in his follow-up reply.

His account names three added inputs in an accuracy follow-up:

• kitchen-island size
• distance from the wall
• countertop height

LinusEkenstam put the Blender modeling run at about 11 minutes and the web-app build at another eight in his runtime breakdown. He later estimated the whole job at 3 to 4% of his monthly quota and less than $10, following a quota reset in a cost estimate.

A similar office demo from higgsfield_ai breaks the handoff into four steps:

1. analyze the reference photos
2. infer missing details
3. build the room in Blender through the Higgsfield plugin
4. publish the model online

LinusEkenstam describes MCP as the main route when it is installed, with computer use handling UI-heavy jobs such as navigating panels and creating cameras in his MCP reply. That distinction explains why the demo produces a real Blender scene rather than a rendered approximation alone.

The model made mistakes in the stairs and mezzanine depth, according to LinusEkenstam's accuracy report. LinusEkenstam called the result a short run rather than archviz, and said he planned a higher-end test using more detailed layouts and reference visuals in an archviz caveat.

He also separated the 11-minute effort from a GLM-5.3-Flash demo he described as taking more than 12 hours and 100 million tokens in the same comparison.

bennash used a related image-to-scene pipeline on M.C. Escher's Waterfall, then turned the result into a stylized camera pass. His prompt sequence laid out the chain:

1. give Astra the source image and ask it to build the structure
2. add an arena and background geometry for every camera angle
3. render a gray-clay fly-through
4. divide the video into three segments
5. generate a textured reference image in an 1980s fantasy-film style
6. restyle each segment in MiniMax H3

bennash said the full process took less than an hour. In another Astra and H3 Max test, gokayfem reported three 15-second 1080p clips in four styles, rendered in under a minute inside Blender in an H3 Max run.

LinusEkenstam says he has also used Astra to model parts for a broken toilet, revise Gridfinity storage, fit containers around objects, and build a browser game about making sandcastles in his object-work reply. Those experiments extend the same reference-plus-constraint pattern from a room-scale scene to functional objects and small interactive projects.