A promising engineering research direction
In Skyfall’s tests, SKY-001 has slightly higher CAD IoU and higher BIM IoU, with much shorter reported generation times and lower task costs. Those figures support a trial, not a production guarantee.
CAD / BIM · Official benchmark analysis
SKY-001 vs GPT-6 Astra: CAD/BIM performance, speed and cost
Compare a specialized engineering research model with a general-purpose reasoning model. Read the reported numbers alongside their units, access requirements and delivery limitations.
Independent analysis of vendor-reported results. We did not run these benchmarks, and this page does not provide model inference or collect waitlist applications.
Sources checked:
What the numbers suggest
Evaluate the integration you can use today
Astra has public API documentation and pricing for broad reasoning, coding and tool-assisted workflows. Account access and limits still apply; CAD/BIM deliverables need validation in the target software.
Different model roles
SKY-001 · engineering world model
Skyfall describes a pipeline that decomposes a design into a hierarchy and predicts visual construction states. The report examines CAD components and BIM floor plans. Assess its outputs against your required formats and editing workflow.
GPT-6 Astra · general-purpose reasoning
OpenAI documents Astra for reasoning, coding, research and tool use. Generating scripts or coordinating software tools is a different integration path from a specialized visual construction model.
CAD and BIM benchmark results
Source: Skyfall Research, 19 October 2026. Both models below are compared in Skyfall’s tests. These numbers have not been independently reproduced by this website.
| Metric | SKY-001 | GPT-6 Astra |
|---|---|---|
| CAD · mean IoU | 79.7% | 79.1% |
| CAD · generation time per case | 3.9 s | 199 s |
| CAD · benchmark cost per case | $0.005 | $0.75 |
| BIM · mean IoU | 43.6% | 40.6% |
| BIM · generation time per case | 1.22 s | 144.6 s |
| BIM · benchmark cost per case | $0.0024 | $0.446 |
How to read the benchmark
CAD: components, not complete production projects
The report uses 1,000 CAD components, with roughly 100 drawing steps per component. Final geometry is compared with a reference using IoU. Do not transfer the per-component timing directly to a whole assembly or an interactive editing session.
BIM: floor-plan overlap
The 1,000 plans include about 350 whole floors and 650 apartment plans, averaging roughly 70 drawing steps. IoU measures overlap; it does not certify dimensions, structural performance, manufacturability or building-code compliance.
From the reported scores, SKY-001 leads by 0.6 percentage points in CAD and 3.0 percentage points in BIM. These are percentage-point differences, not relative improvements in engineering correctness.
Reproduction limits: we have not verified complete model snapshots, reasoning settings, hardware, retry policy or end-to-end pipeline coverage for every comparison. Reported generation times are not API latency commitments or service-level guarantees.
Benchmark task cost vs API pricing
The benchmark table above measures USD per test case. The API table below uses USD per million text tokens. Different units and workloads prevent a direct retail-price comparison.
| Billing item | GPT-6 Astra · public rate | SKY-001 · public rate |
|---|---|---|
| Uncached input | $10.00 | Public quote not yet verified |
| Output | $50.00 | Public quote not yet verified |
| Cached input | $1.00 | Public quote not yet verified |
The Astra rates shown apply to Standard short-context text usage. Inputs above 272K tokens trigger higher rates for the full request: 2× input/cache rates and 1.5× output rates. Cache writes and tools can add charges; check the current rate card, cache TTL and your actual usage.
SKY-001’s reported $0.005/CAD case and $0.0024/BIM case are experimental task costs, not a verified customer tariff. We have not verified a public SKY-001 commercial rate card or self-service API contract. Ask the team about access and billing.
Budget for accepted deliverables: include generation, retries, tool execution and human correction, then divide by the number of outputs that pass your acceptance checks. Token rates or a single benchmark case alone do not establish that cost.
Access and delivery boundaries
SKY-001: verify access with the team
The official product page currently directs visitors to a waitlist. Confirm availability, supported inputs and outputs, export formats, editing capabilities and commercial terms before planning an integration.
Astra: public API documentation
Consult the model documentation and check your account’s access, limits and tool support. A documented API does not itself provide an editable CAD/BIM deliverable or a turnkey engineering integration.
In the report’s “What Comes Next” section, executable CAD operations with editable feature history and editable BIM building elements are future directions. This guide does not treat those roadmap items as current delivered capabilities.
Which model should you evaluate?
Engineering design exploration
Consider applying to trial SKY-001 for repeated component or layout exploration. Validate actual access, output format and quality with representative tasks before relying on benchmark cost or speed advantages.
Coding and cross-task workflows
Evaluate Astra when you need a documented API for code, reasoning and software-tool coordination. Measure the complete CAD/BIM workflow, including script execution, validation, retries and human edits.
Editable CAD/BIM deliverables
For either model, test file compatibility, dimensions, constraints, editable history and manual rework in the target application. Choose only after defining acceptance criteria for your own engineering task.
Frequently asked questions
Official sources and verification date
Use the original report for methodology and the providers’ current pages for access and pricing. This is an independent resource with no affiliation to Skyfall AI or OpenAI.