SKY-001 vs GPT-6 Astra: CAD and BIM Claims

SKY-001 and GPT-6 Astra solve different layers of the problem.

On the official SKY-001 page, Skyfall describes a hierarchical visual world model that explores a design, decomposes it into parts, draws construction states and assembles the result. Its first public focus is CAD and BIM. Access is described as a research preview with a waitlist.

OpenAI describes GPT-6 Astra as a broad model for computer use, professional work, software engineering and science. The release includes BenchCAD among its evaluations, but Astra is not presented as a CAD-only model.

SKY-001

Specialized engineering world model; public focus on CAD and BIM; waitlist access.

GPT-6 Astra

General professional model; broader tool use, code, research and design tasks.

Direct comparison

Possible on a fixed engineering task, but only with the same inputs, outputs and scoring rules.

What does Skyfall claim in its SKY-001 comparison?

Skyfall’s launch page reports the following results against GPT-6 Astra. These are reproduced here as claims from the model provider; AdverForge has not run the benchmark and the figures are not independent findings.

IoU measures geometric overlap with a reference. It does not by itself prove that a design meets dimensions, constraints, building codes, manufacturing requirements or human engineering intent. Speed and cost also depend on the complete harness, compute and retry policy.

Which model fits which job?

• Evaluate SKY-001 when the task is explicitly about step-by-step geometry, CAD construction or BIM floor-plan generation and research-preview access is acceptable.
• Evaluate GPT-6 Astra when the workflow also needs research, code, browser or computer use, document work and coordination across several tools.
• Use an established CAD or BIM system as the final authority. Neither a visually plausible rendering nor an IoU score replaces design constraints, standards review and an engineer’s approval.

There is not enough independent evidence here to call one model the universal winner. The correct choice depends on the deliverable and how much of the engineering workflow must remain editable and verifiable.

How to run a fair SKY-001 vs GPT-6 Astra test

1. Freeze the input. Use the same prompt, image, units, constraints and target file format.
2. Define acceptance before running. Score geometry, dimensions, constraint validity, editability and completion rate—not appearance alone.
3. Record the full setup. Include model version, tool harness, compute, retries, wall time and total billed cost.
4. Review failures. Count missing parts, invalid topology, impossible construction steps and outputs that require manual repair.
5. Use an independent reviewer. Keep the evaluator blind to the model name and publish enough detail for someone else to repeat the test.

What does this comparison mean for AI video?

Neither model is the video renderer used by AdverForge. An engineering model could help describe a product assembly or prepare a structured design brief, but producing product footage is a separate task. AdverForge uses Seedance for video generation and does not integrate SKY-001 or GPT-6 Astra as a generation fallback.

If you want to turn an approved product reference into footage, see the Seedance 2.5 product-video guide and video templates with playable examples.

Sources and scope

Product descriptions and benchmark figures were checked against Skyfall’s official SKY-001 page and OpenAI’s official GPT-6 Astra release on October 21, 2026. Benchmark interpretation and the test plan are AdverForge editorial analysis. We have not independently benchmarked either model and are not affiliated with Skyfall or OpenAI.

Frequently asked questions