Skyfall AI's SKY-001 arrived on October 19, 2026 with a headline that invites comparison: the same CAD quality as GPT-6 Astra at about 150x lower cost. But "world model" now covers robot brains, video simulators and design planners, and putting them in one leaderboard is how people end up comparing the wrong things.
This post makes three comparisons that are actually fair. First, SKY-001 against the frontier language models Skyfall tested. Second, SKY-001 against the cheaper and weaker frontier options. Third, SKY-001 against the other world models we have covered, where the honest answer is that they do different jobs. For the launch details, read our SKY-001 breakdown first.
TL;DR: who should use which
| Question | Short answer |
|---|---|
| Best CAD quality per dollar (per Skyfall's data) | SKY-001: 79.7% IoU at $0.005 per case |
| Best CAD you can use today | A general frontier LLM with a CAD tool; SKY-001 is waitlist-only |
| Closest frontier rival on quality | GPT-6 Astra at 79.1% IoU, then Gemini 4 at 76.9% |
| Best for controlling a robot | Dream-1 (public API, $3 per hour) or GreenWAM and Long-WAM if you want open research |
| Best for simulating physical scenes in video | Odyssey-3, which reports 66.1 on Physics-IQ Verified |
| Biggest caveat on SKY-001 numbers | Company-run benchmark, small accuracy lead, no independent replication |
Why is comparing world models so confusing?
The label covers at least three different things. A design-and-build planner like SKY-001 predicts which parts a product has and the ordered steps that make each one. A world-action model like GreenWAM, Long-WAM or Dream-1 predicts how a robot's scene changes and which action to take next. A video world model like Odyssey-3 predicts how a scene evolves frame by frame so you can explore or simulate it.
Each is scored differently. SKY-001 uses IoU, the overlap between a generated shape and a reference. Robot models use task success on suites such as RoboCasa, RoboTwin and LIBERO. Video models use physics-prediction scores. A 79.7% IoU and a 78.7% task success rate are not on the same ruler. If you want the category basics, see our explainer on what world models are.
Two diverging paths, a picture of choosing between a frontier language model and a specialist world model for the same engineering job
Comparison 1: SKY-001 vs GPT-6 Astra
GPT-6 Astra is the baseline Skyfall chose, and with reason: Skyfall's own post calls it widely recognised as the state of the art for this task. It is also the model OpenAI positions as its top tier. For pricing context on the GPT-6 family, see our GPT-6.1 Sol and Astra cost-efficiency breakdown.
| Metric | SKY-001 | GPT-6 Astra | Difference |
|---|---|---|---|
| CAD IoU (1,000 components) | 79.7% | 79.1% | +0.6 points |
| CAD time per case | 3.9 s | 199 s | about 51x faster |
| CAD cost per case | $0.005 | $0.75 | about 150x cheaper |
| BIM IoU (1,000 plans) | 43.6% | 40.6% | +3.0 points |
| BIM time per case | 1.22 s | 144.6 s | about 119x faster |
| BIM cost per case | $0.0024 | $0.446 | about 186x cheaper |
| Composer judge score, part decomposition | 7.8 | 6.4 | +1.4 on a 10-point scale |
All figures come from Skyfall's launch post. The composer scores were produced by two blind LLM judges, GPT-6.1 Sol and Claude Opus 5.5, and Skyfall notes GPT-6 Astra returned a model for only 82% of assets.
What this comparison gets right
The method difference is real. GPT-6 Astra writes CAD or Blender code in one request, which is a long token-by-token chain. SKY-001 plans a tree of parts and draws visual steps, so the sequential depth is small and the pieces run in parallel. That structural advantage explains why a small model can be fast.
Where to be careful
- A 0.6-point IoU edge on CAD is a tie in practice. Treat the CAD result as "equal quality."
- The prompts, scripts and retry policy used for GPT-6 Astra are not public. A different harness can move a frontier model's score a lot.
- The cost number for the LLM includes reasoning tokens. A cheaper effort setting might change it.
- BIM at 43.6% is still a low absolute score. A 3-point lead on a hard task is not the same as solving it.
Verdict: if Skyfall's numbers replicate, SKY-001 wins on economics and ties on CAD quality. Until then, GPT-6 Astra remains the model you can actually call.
Comparison 2: SKY-001 vs Gemini 4, Qwen, Kimi and Grok
The more useful table for a team on a budget is the whole field. Skyfall tested five frontier models on CAD and, for BIM, a slightly different set.
| CAD model | IoU | Time | Cost |
|---|---|---|---|
| SKY-001 | 79.7% | 3.9 s | $0.005 |
| GPT-6 Astra | 79.1% | 199 s | $0.75 |
| Gemini 4 | 76.9% | 253 s | $0.25 |
| Qwen 3.8 Max | 34.4% | 764 s | $0.30 |
| Kimi K3 | 9.5% | 439 s | $0.13 |
| Grok 4.7 | no output | 1,250 s | $0.59 |
| BIM model | IoU | Time | Cost |
|---|---|---|---|
| SKY-001 | 43.6% | 1.22 s | $0.0024 |
| GPT-6 Astra | 40.6% | 144.6 s | $0.446 |
| Gemini 3.8 Flash | 42.3% | 346.2 s | $0.218 |
| Qwen 3.8 Max | 28.8% | 255.3 s | $0.111 |
| Kimi K3 | 33.7% | 113.9 s | $0.270 |
| Grok 4.7 | 28.2% | 592.1 s | $0.305 |
Two model cards side by side, a picture of comparing SKY-001 and Gemini 4 on speed, cost and CAD accuracy
Reading the table
Gemini 4 is the interesting second place on CAD. It sits 2.8 points behind SKY-001 and costs $0.25 per case, so the cost gap to SKY-001 is about 50x rather than 150x. If you need a callable model today and can accept the lower accuracy, it is the value pick among the LLMs. Our Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra comparison covers how they differ outside this benchmark.
Qwen 3.8 Max and Kimi K3 are cheap but fall off a cliff on CAD, at 34.4% and 9.5% IoU. The low cost does not help if the shape is wrong. On BIM they recover to 28.8% and 33.7%, because drawing a floor plan is an easier output than writing a long CAD program.
Grok 4.7 produced no CAD output at all, and Skyfall notes it completed none of the cases. On BIM it scored 28.2%. Read that as "fails the long program," not "bad model."
One more point: Claude Opus 5.5 appears in Skyfall's X thread as a comparison and as a judge, but not in the printed CAD and BIM tables. We could not see a Claude row to compare against, so we do not rank it here. For its general strengths see the Opus 5.5 launch and benchmarks.
Verdict: among models you can call now, GPT-6 Astra leads on CAD accuracy and Gemini 4 on value. SKY-001 leads both if the data holds.
Comparison 3: SKY-001 vs GreenWAM, Long-WAM, Dream-1 and Odyssey-3
This is the one where a head-to-head score would mislead. The table compares jobs, not leaderboards.
| Model | What it predicts | Output | Headline number | Access |
|---|---|---|---|---|
| SKY-001 | How a design decomposes and is built | Part tree, CAD sketches, floor-plan states | 79.7% CAD IoU, 43.6% BIM IoU | Waitlist |
| GreenWAM | How a robot scene changes and which action to take | Actions and predicted scenes for 20 robot families | 67.54 mean on RoboTwin 2.0 | Open weights on Hugging Face |
| Long-WAM | Robot actions with long video history | Action chunks at 107.4 ms | 78.7% on RoboCasa GR-1 at 19.2 s context | Research paper and project page |
| Dream-1 | Actions for unseen two-armed robots | End-effector poses | Zero-shot on 27 unseen robots in simulation | API, $3 per hour |
| Odyssey-3 | How a scene evolves over time | Video frames | 66.1 on Physics-IQ Verified (Pro, best-of-8) | Free research preview, API by contact |
What they share
All of these models predict visual states instead of text tokens, and most of the recent gains come from breaking a long horizon into short, structured steps. Long-WAM does it with causal video history. SKY-001 does it with a hierarchy of parts. GreenWAM does it by sharing an action space across 20 robot bodies.
What separates them
- SKY-001 produces a plan, not behavior. Its output is a design and its build order. Nothing in the launch post says it controls a machine.
- The robot models act on the world. GreenWAM, Long-WAM and Dream-1 close a loop with a physical robot, so latency and safety matter more than they do for a design tool.
- Odyssey-3 simulates. It is closer to a physics-aware video generator, and its own leaderboard has caveats such as best-of-8 sampling and cost figures that exclude prompt-rewriting fees.
A realistic stack could chain them. A design model proposes a part, and a robot model picks it up. No vendor has shown that pipeline, so treat it as an idea, not a product. Our earlier piece on Dyna-2 and world-action model scaling covers how the robot side is scaling.
Verdict: do not choose between SKY-001 and these models. Pick by the job: design (SKY-001), robot control (Dream-1, GreenWAM, Long-WAM), or physical scene simulation (Odyssey-3).
What people are asking
Can I replace my CAD agent with SKY-001? Not yet. It is waitlist-only, and Skyfall lists editable feature histories and executable operations as future work. Today's open path is an agent driving a parametric kernel, as in our guides to SolidCraft and the text-to-CAD agent plugin.
Is 150x cheaper believable? The arithmetic matches Skyfall's table, and a small specialized model running a few seconds on a GPU being far cheaper than a large model reasoning for 200 seconds is plausible. What we cannot judge is the hardware cost basis behind the $0.005 figure.
Does this threaten the frontier labs? It is a narrow benchmark on two task families. The more credible reading is that specialists can undercut generalists on well-defined visual tasks, the same argument we covered for Liquid AI's D1 model.
Why does IoU matter? IoU is the overlap between the generated and reference shape, so it rewards geometry, not manufacturability. A part can score well and still be impossible to machine. Skyfall does not report manufacturability checks.
How should you decide?
- You need CAD output this week: use a frontier LLM with a parametric CAD tool. GPT-6 Astra for accuracy, Gemini 4 for budget.
- You run thousands of design variations: join the SKY-001 waitlist and ask for a trial on your own parts. The cost gap only matters at volume.
- You work on robots: compare Dream-1's hosted price with GreenWAM's open weights. Ignore SKY-001.
- You need physics-aware video: look at Odyssey-3 and read its caveats first.
- You are choosing for a roadmap: wait for an independent benchmark of SKY-001 before committing.
Related reading
- Skyfall SKY-001: engineering world model, claimed vs verified
- What are world models? Starchild-1 and Odyssey explained
- GreenWAM: open robot world model for 20 robot families
- Long-WAM: NVIDIA world-action model
- Dream-1 robot world-action model API
- Odyssey-3 Physics-IQ Verified results
- Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
- Official: Skyfall's SKY-001 research post
Benchmark figures for SKY-001 and the comparison models come from Skyfall AI's October 19, 2026 launch post and are the company's own results. Other models' figures come from the linked explainx.ai posts. All are accurate as of the publication date and may change as independent tests appear.