Key Takeaways
AI can now work with editable geometry through existing CAD and Modelling tools. Astra writes parametric programs, Fable executes long workflows, and Atlas captures physical environments in point clouds.
A step change in a narrow domain needs two things: the data, and a harness that can act on it. AEC holds both.
The durable capability is designing the operating loop around the model. A proper loop needs three layers: specs, verification, and the environment.
BIM can become AEC’s verification layer. Schemas, clash rules, quantities, and model checks can determine whether an output is acceptable and allow for AI to perfect it.
For thirty years the assumption was that AI would reach the construction industry through documents: specifications, RFIs, contracts, submittals. It did, with mixed success. But the announcements of early September 2026 point somewhere else, and they promise the moment the Architecture, Engineering and Construction (AEC) industry has been preparing for.
This one is about geometry. And it arrives at an industry that already owns what other are now scrambling to build. One industry after another is now building a verification layer from scratch. AEC spent thirty years building one and called it BIM.
On 1 September 2026, World Labs announced Atlas, an “omni world model” pretrained from scratch to operate natively on text, images, video and 3D, rebuilding a real environment from as few as two or three images, and exporting point clouds or 3D Gaussian splats [1]. Two days later, on 3 September, OpenAI shipped GPT-6 Astra, and for the first time a flagship general model’s launch material led with CAD and 3D work rather than text and code [2]. Alongside them, also on 1 September, Anthropic shipped Claude Fable 5.1, which made no 3D claim whatsoever. Its gains were in long-horizon agentic work [3].
In this post, I aim to review the following
Astra, Fable, and Atlas
OpenAI GPT 6 Astra in 3D modeling
Anthropic Fable 5.1 in BIM, scheduling, and shape files
World Labs Atlas in point clouds
Why this might be AEC’s moment
Change will come in major steps
AI development triad: spec, verification, and the environment
BIM as a verification layer
Sector specific AI developments
Human-AI collaboration mindset
Question then becomes
Let us dive in.
1. Astra, Fable, and Atlas
Most coverage of the GPT-6 Astra launch described a model that builds 3D models, on the strength of a reported 95.9% on the BenchCAD leaderboard. The details behind that number matter.
BenchCAD is the benchmark behind a test of 17,900 CAD programs across 106 industrial part families, roughly half anchored to published ISO, DIN, EN, ASME or IEC standards, every one compiled and run before it entered the set [4].
What it measures is narrow and real: does the model’s code execute, and does the shape it produces occupy the same space as the reference answer. What it does not measure is everything around the shape, including manufacturing tolerances, undocumented design intent, assembly information, and whether a part complies with the standard it was anchored to rather than merely resembling it [4].
A high score means a model can write a program scripts that runs in a modelling software to produce a geometric model on a test the vendor administered to itself. It does not mean the model understood why the part is shaped that way [5].
Independent CAD reviewers report models getting the broad shape right while failing parametric intent: missing detail, misreading engineering parameters, choosing the wrong modelling steps [6].
1.a. OpenAI GPT 6 Astra in 3D modeling
The demonstrations read more clearly than the scores. GPT-6 Astra drives Blender (3D modelling software) from a plain text prompt and prepares assets for Unreal Engine 5, and the game studio Playco built three themed playable prototypes from a single grey-box model, reporting 50% fewer manual fixes [7].
An interesting public example is the reconstruction of the Palace of Fine Arts, a full Blender scene built from reference imagery in one session [8].
![GPT-6 Astra building a full 3D scene of the Palace of Fine Arts in Blender: the assets and textures the model assembled, above, and the rendered result, below. Public demonstration by Sharif Shameem [8].](https://substackcdn.com/image/fetch/$s_!tdRc!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc0ee56c6-d978-4e13-b10f-c9db09745eca_936x497.png "GPT-6 Astra building a full 3D scene of the Palace of Fine Arts in Blender: the assets and textures the model assembled, above, and the rendered result, below. Public demonstration by Sharif Shameem [8].")
The sequence below is roughly the process that the model followed.
Read the reference into a scene description. Astra takes the photographs as input and writes down what is in them: a domed rotunda, a colonnade, a peristyle wrapped around a lagoon, and the massing relationships between the three. Language, not pixels.
Fix the geometry as parameters. Rotunda radius, column diameter, bay spacing, cornice heights and lagoon extents become named values, so the scene can be rebuilt at a different scale without being redrawn.
Write the generator, not the model. The output is Blender Python. The walkway is a loop over a bay spacing rather than a row of individually placed columns, which is why the finished scene arrives as editable objects instead of one mesh.
Run it, read the error, edit the code. The script executes inside Blender and the model reads back what happened, whether a traceback or a scene that does not match what it asked for, and rewrites its own program. This is the long-horizon loop, and it is where most of the session goes.
Layer materials, texture and environment. Surfaces, the water plane and the surrounding planting are added as separate passes over the geometry the script has already placed.
Set a matching camera and render. A view is framed to match the reference photograph so the render and the source can be put side by side. It is the only check in the whole sequence performed by eye.
The third step is the one that matters. What Astra produced is a program that builds the Palace of Fine Arts, not a picture of it. Change the bay spacing and the walkway rebuilds. That is a different kind of artefact from a mesh.
1.b. Anthropic Fable 5.1 in BIM, scheduling, and shape files
Claude Fable 5.1 makes no CAD claim of its own. Anthropic’s launch material published no BenchCAD figure, but the 84.3% is its score on the BenchCAD leaderboard on a self-reported basis as Astra’s [5].
That the number is respectable anyway, and last only tow days at top of the leaderboard is worth a moment’s thought.
Complex modelling is rarely one act of geometric insight; it is a long chain of small ordinary decisions: pick the sketch plane, extrude, fillet, check the result, notice the wall went the wrong way, undo, try again. Parametric CAD punishes the first wrong step and rewards patience, so a model that can carry a set of constraints through a fifty-step edit session and recover from its own mistakes will out-model one that writes a more elegant program in a single pass.
Claude Fable 5.1’s primary reported gains sit in exactly that type of work: long-horizon agentic tasks where a model holds a goal across many steps, calls tools, reads what came back, and corrects itself rather than committing to its first answer [3].
Two demonstrations, I believe, could show what that looks like in our own work. Both were run on Claude Fable 5.1, each in a single session, with no human intervention between the opening instruction and the final file.
Photo → BIM
Asking Claude to create a BIM model from a picture. One street-level photograph in, an IFC4 model at LOD 200 out. And the model took the following seven steps:
Read the photo into primitives. Volumes, roof planes, openings, cladding zones, a scene description, not pixels.
Scale off standard components. E.g., Garage door 16 × 7 ft.
Fix a frame, parameterize. X/Y/Z convention; every dimension a named constant with its basis.
Build the IFC skeleton. Project → site → building → two storeys, metric, Body context.
Generate the elements. Swept profiles, not meshes, editable in native BIM tools.
Add openings, materials, provenance. Void/fill relationships, surface styles, confidence on every element.
Verify. Schema check, tessellate all 40 products, render against the photo. Caught the garage roof pitch.
BIM → Schedule
Asking Claude to create a construction schedule from the BIM model. The schedule was based on the geometry that was already in the model. Three steps:
Extract the inventory. GlobalIds, classes, volumes, the quantity basis and the join key.
Write the schedule as data. 70 activities, logic, crews, costs, each keyed to the elements it builds.
Run CPM, emit XER, verify. Working-day passes → 17 P6 tables → re-read with two parsers. Caught two dangling activities.
The last step of each run is the one worth noticing. The model checked its own output and found its own mistakes, a wrong garage roof pitch in the first, two dangling activities in the second, and neither was caught by a person. That is long-horizon behaviour doing work an engineer would otherwise do.
Calgary Rail Crossing Corridor
For the third example, I asked Claude Fable 5.1 to pull open geospatial data, built the shapefiles itself, and rendered the same rail network at different scales: the four subdivisions that meet at Alyth Yard, and the downtown corridor in detail.
That matters because scale is where most project geometry gets lost, since the network model, the corridor model and the structure model are usually three separate exercises. A system that can assemble the data and move between those scales inside one session is doing the work of coordination, not just the work of drafting.
1.c. World Labs Atlas in point clouds
World Labs’ Atlas does not appear on the same scale as is a different kind of system. It accepts text as a prompt but does not respond in text. The model reconstructs 3D scenes from as few as two or three images and exports point clouds or 3D Gaussian splats [1].
Atlas emits geometry directly, so a code-execution benchmark like BenchCAD has nothing to say about it. Reconstruction can be measured against ground truth, and World Labs reports a mean absolute-relative pointmap error of 25.3 against 28.7 for the next best system, a vendor-run result that has not been independently replicated [9].
Almost every project begins with a condition somebody has to measure, and today that means a scanner, a crew, a registration workflow in only few days. That is why existing-conditions models are skipped on small jobs and stale on large ones.
Point clouds and splats are not BIM, and nothing here authors a wall or a schedule. But as a source of geometric truth to check a model against, progressive capture is the input our clash, quantity and progress machinery has always lacked.
Read together, GPT-6 Astra, Claude Fable 5.1 and Atlas are not competitors on one axis. They are different answers to how a machine touches three-dimensional space: Astra writes the program, Fable runs the long errand, Atlas perceives the scene.
2. Why this might be AEC’s moment
2.a. Change will come in major steps
Capability in a narrow domain rarely improves along a smooth curve. It sits flat, and then it steps.
Language models playing chess is the clearest worked example. Performance across models and versions was erratic enough to look like a puzzle about reasoning, and the closest examination available traced most of it to two ordinary things: what was in the training data, and how the model was prompted and allowed to act [11]. Neither is a claim about intelligence and both are decisions someone made.
OpenAI has said nothing public about GPT-6 Astra’s training data composition, so what follows is inference rather than fact. But the same two levers are the plausible ones.
The first is data: a deliberate injection of parametric CAD programs, Blender and Grasshopper scripts, IFC and USD structures, and 3D workflow traces into the training corpus.
The second is the harness: computer use maturing enough to let latent spatial competence act. The model may always have half-known Blender; what changed is that it can now click.
The two are not alternatives. Data without a harness is knowledge the model cannot apply, and a harness without data is a model clicking through software it does not understand. A step change in a narrow domain generally needs both.
AEC owns an enormous body of structured, semantically rich, geometrically precise three-dimensional data: thirty years of coordinated federated models, as-builts, clash histories and fabrication models, much of it schema-conformant, and most of it authored by people who were paid and professionally liable to get it right.
2.b. The AI development triad: spec, verification, and the environment
One useful way to understand AI deployment is to divide it into three connected layers: write down what the tool must do, decide in advance how you will test that it does it, and build the surroundings it runs in. That is Spec, Verification, and Enviroment.
The division is not new, which is the strongest argument for it. Anyone who has specified engineering software has worked in a relatively similar way.
The layers are connected rather than sequential: a weak spec makes a verifier impossible to write, and a poor environment makes a good spec unenforceable.
Spec: a precise statement of what “done” means for a task, explicit enough that an agent does not have to infer the intent. Specs are built iteratively through use, not dumped upfront in a single document.
Verification: an independent and ideally objective test of whether an output meets its spec. It answers accept or reject without a human forming an opinion. “Does it look right” is not a verifier.
Environment: the persistent workshop an agent operates in, holding standing instructions, accumulated project knowledge, reusable skills and guardrails around destructive actions. It is what makes agent behaviour repeatable across tasks.
For complex domains and tasks, verification is often the bottleneck.
For AI development too, capability is no longer the binding constraint; verification is.
2.3 BIM as a verification layer
Our industry appear to have built the perfect verification layer in past 30 years.
A coordinated building information model is not a drawing, and it is not a rendering. It is a structured, queryable, rule-checkable assertion about a building, carrying clash detection, model-checking rule sets, schema validation against the Industry Foundation Classes standard [13], quantity extraction, and a research literature on automated rule-based code compliance that predates the current wave of models by well over a decade [14]. Each of those is a verifier in the technical sense.
One industry after another is now building verification layers from scratch.
AEC spent thirty years building one called BIM.
This is why the text-only detail from Astras process matters. GPT-6 Astra’s contribution to AEC is not that it makes geometry. It is that it emits executable, checkable artefacts. Pair an agent that emits checkable artefacts with an industry that already owns the checker, and the loop closes.
If BIM is the verification layer, the next question is whether the sector is ready to use it as one. One thing for sure, the product activity is real: three AI-BIM products launched in June 2026 alone, Trunk Tools Cortex, Beam AI’s BIM CoPilot and OFA Group’s QikBIM, alongside Higharc’s US$95M Series C, which is a funding round rather than a product launch [15].
3. Sector specific AI developments
In March 2026 Anthropic published its own read of the labour market, and one finding lands directly on us [16]. On 10 September 2026, OpenAI launched ChatGPT for Financial Services. It is built on GPT-6 Astra, with Morgan Stanley and Evercore as design partners. It connects to PitchBook, S&P Capital IQ, LSEG and Moody’s, and it produces research, models and pitchbooks in a firm’s own templates [23].
The two developments may shine some light on potential trajectories of AI in AEC.
The Anthropic report introduces a simple idea. Earlier studies asked what AI could do in an occupation. Anthropic asked what it is actually being asked to do. The first they call theoretical exposure. The second they call observed exposure, and they build it from real usage rather than from estimates [16].
For architecture and engineering the two numbers sit further apart than for any other major occupational group. Dezeen reports roughly 70% theoretical exposure against observed exposure close to zero; Monograph reports about 85% theoretical against about 5% observed [17,18]. The table below shows the numbers for business and finance, 94.3% theoretical versus 28.4% observed.
Frontier labs are going vertical, and the playbook is the same each time: pick a sector with high exposure, connect the model to the data and templates that sector already trusts, and build it with the firms that set its standard of work.
AEC fits that profile better than most. We have the widest exposure gap of any major occupational group, a mature data layer in BIM and IFC, and now a flagship model whose launch material led with CAD [2]
4. Human-AI collaboration mindset
The collaboration between Human and AI is still in its infancy, and while it is, it pays to watch how others use it.
Reading what power users and whole organizations actually do with these tools is consistently more useful than reading what the vendors say they do, and the first thing you notice is how differently everyone works.
Anthropic’s own usage study makes the point at scale: what people ask for shifts by hour of the day, by day of the week and by occupation, and personal use climbs from about 35% of conversations on a weekday to roughly 50% at the weekend [19]. There is no single way to use one of these systems.
Commonalities do emerge, though, and they are the interesting part. The people getting the most out of these tools are not writing better prompts. They are building something that runs.
Which is how the vocabulary changed. Boris Cherny, who leads Claude Code at Anthropic, put it about as plainly as it can be put in June 2026: “I don’t prompt Claude anymore. I write loops and the loops do the work. My job is to write loops” [20].
The label matters less than what it concedes, which is that the leverage has moved from the model to the structure around it.
A spec, a verification and an environment are a loop.
A second idea explains why the loop has to be built rather than assumed. Andrej Karpathy draws a distinction between animals and ghosts [21]. An animal arrives with drives it did not learn. It is hungry, it is curious, it wants things and pursues them without being asked. A large model has none of that. It is closer to a ghost summoned out of a corpus of human text, capable and inert at the same time, with no intrinsic interest in whether your drawing set comes out right. A system with no goals of its own will not notice that it has drifted, will not mind that the model no longer matches the specification, and will not stop. (This goes to the core of the alignment debate, which is a topic for another day).
A third idea is Ethan Mollick’s, and it is about timing rather than architecture: the AI you are using today is the worst AI you will ever use [22].
The model is the part that improves without your help, so building your work around this quarter’s model capability means building on the thing that will change, while building the loop around it means building on the thing that will last. A verifier written against IFC outputs this year runs unmodified against next year’s model, and returns better results, because the thing in the middle of the loop got stronger.
A good human-AI collaboration mindset is aiming to build the best loops for the worst AI you will ever use.
5. The question then becomes
With all the buzz about AI tools, perhaps there is solace in the long, struggling hours it takes to shape their results into a useful artefact, such as this letter. Those hours invite deeper thoughts and questions about how we should react to this new invention.
One major thought I had is this: if verification makes autonomy possible, could it also become a ceiling? If we delegate only what can be measured, will we eventually value what is checkable more than what is important? And when that happens, what will be important?
If AI performs the repetitive work through which young professionals once developed intuition, can the industry preserve expert judgment without preserving the experiences that created it?
Our codes and verification rules preserve lessons from the past. When AI proposes something genuinely new, should the verifier protect society from novelty, or protect precedent from being challenged?
Please share your thoughts.
Notes
Join the EPM Network to access insights, influence our research, and connect with a community shaping the industry’s future.
Support us by sharing this article with your friends and colleagues, or over social media.
If you wish to share your opinion, provide insights, correct any details in this article, or if you have any questions, please email editor@epmresearch.com.
Refer to this article using the following citation format:
Zangeneh, P. (2026), “AI Takes Shape: What Recent AI Advances Mean for Architecture, Engineering, and Construction,” EPM Research, 11 September 2026.
Partner with EPM Research
EPM Research at the University of Calgary aims to be an effective research arm for Canadian Project and Construction Management industry.
Generative AI in construction and project management is a core track of our research. Our work ranges from AI adoption strategy to independent reviews of AI systems to pilot development and implementation. We place a particular focus on supporting small and medium-sized Canadian organizations in the public and private sectors, a priority encouraged by the Engineering Project Management Endowment and Canadian research funding agencies.
Please use this form to tell us what you're working on and how we might help. We review submissions regularly and will contact you when there is a good fit with our research directions and capacity.
About the author
Pouya Zangeneh, Ph.D., P.Eng., is Assistant Professor in the Department of Civil Engineering and Endowed Chair in Engineering Project Management at the Schulich School of Engineering, University of Calgary. He has more than 15 years of experience in multi-billion-dollar capital projects and capital investment planning across heavy industrial, utilities, energy, mining, and infrastructure sectors.
Related reading from EPM Research
References
World Labs. (2026, September 1). Atlas: A world model for spatial intelligence. World Labs.
OpenAI. (2026). GPT-6 Astra: A new generation of intelligence. OpenAI.
Anthropic. (2026, September 1). Introducing Claude Fable 5.1 and Claude Mythos 5.1. Anthropic. With model specifications from llm-stats (2026) and CellCog (2026), Fable 5.1 pricing, context window and benchmarks. https://www.anthropic.com/claude-fable-and-mythos-5-1
BenchCAD: A comprehensive, industry-standard benchmark for programmatic CAD. (2026). arXiv:2605.10865.
BenchCAD. (2026). Official leaderboard, benchcad.com/leaderboard, including self-reported, with-tools entries flagged as not re-graded; and llm-stats. (2026). BenchCAD leaderboard mirror, which reports a differing set of figures. https://llm-stats.com/benchmarks/benchcad
NoriCAD. (2026). GPT-6 Astra and CAD: Could AI change 2D and 3D design?. NoriCAD.
Crypto Briefing. (2026). OpenAI’s GPT-6 Astra enables 3D modeling from Blender to Unreal Engine 5; and OpenAI. (2026). Playco: game prototyping with Astra. OpenAI customer story (source of the customer-reported 50% fewer manual fixes). https://cryptobriefing.com/openai-astra-3d-modeling-unreal-engine/
Shameem, S. (2026, September). Astra 3D reconstruction of the Palace of Fine Arts in Blender. X.
AlphaSignal. (2026). World Labs’ Atlas beats specialized 3D models with one omni model: reported sparse-view reconstruction and camera-control preference results. https://alphasignal.ai/news/world-labs-atlas-beats-specialized-3d-models-with-one-omni-model
OpenAI. (2026). GPT-6 Astra model card. OpenAI API documentation.
Dynomight. (2024). OK, I can partly explain the LLM chess weirdness now. Dynomight.
Wing Venture Capital. (2025). RL environments for agentic AI: Who will win the training and verification layer by 2030. Wing VC.
ISO 16739-1:2024. Industry Foundation Classes (IFC) for data sharing in the construction and facility management industries, Part 1: Data schema. International Organization for Standardization.
Eastman, C., Lee, J., Jeong, Y., & Lee, J. (2009). Automatic rule-based checking of building designs. Automation in Construction, 18(8), 1011–1033. https://doi.org/10.1016/j.autcon.2009.07.002
Geo Week News. (2026). AI BIM automation sees concentrated June activity with four major launches. Geo Week News.
Anthropic. (2026, March 5). Labor market impacts of AI: A new measure and early evidence. Anthropic. (Source of the observed-exposure measure, the 0.6-point BLS growth relationship, and the 22–25 hiring figures.)
Dezeen. (2026, March 11). Architects and engineers among professions most automatable by AI according to Anthropic. (Reports architecture and engineering at roughly 70% theoretical exposure with observed exposure close to zero.)
Monograph. (2026). Why architecture and engineering have the most untapped AI potential of any industry. (Reports the same gap at roughly 85% theoretical against about 5% observed, and the American Institute of Architects 2025 figures: 6% of practising architects using AI regularly, 8% of firms with AI in a process. The two write-ups differ; both are secondary coverage of the same report.)
Anthropic. (2026, June 26). Anthropic Economic Index report: Cadences. Anthropic. (Source of the weekday versus weekend usage split, the 93% artefact figure, and the note that construction and extraction occupations are under-represented.)
The New Stack. (2026, June 10). The Anthropic leader who built Claude Code says he ditched prompting — now he just writes loops. (Source of the Boris Cherny quotation.)
Patel, D. (2025, October). Andrej Karpathy: “We’re summoning ghosts, not building animals.” Dwarkesh Podcast. Primary source for the animals-versus-ghosts distinction and the “ethereal spirit entities” phrasing; see also Karpathy, A. (2025), Animals vs ghosts, and TeamDay (2025), We’re building ghosts, not animals, for secondary write-ups.
Mollick, E. One Useful Thing. Source of the “the worst AI you will ever use” framing, repeated across his writing
Open AI, (2026, September 10). Introducing ChatGPT for Financial Services. OpenAI.