On 29 October, Snaptrude launched Snaptrude Agent, an AI agent for its browser-based BIM platform, and published AECBench, a benchmark for scoring AI on architectural work. The two together are more interesting than either alone, because the benchmark is the company’s answer to the obvious question about the agent — how good is it — and because the company has published the numbers rather than a claim.
What the Agent Does
Per AEC Magazine’s write-up, which carries the fullest capability list, Snaptrude Agent is meant to run from early concept studies through to construction-ready LoD 400 models including structure, mechanical, electrical and plumbing systems, and drawings. The interaction loop is prompt, plan, approve: a user gives the Agent a prompt or a design problem, the Agent proposes a plan, and the user approves it or redirects before it executes.
The specific claims, all Snaptrude’s own as reported there:
- Generate massing options from a site analysis — “from a handful to hundreds.”
- Convert massing models into detailed LoD 400 BIM models in minutes.
- Model structural and MEP systems parametrically.
- Work from vendor specification sheets for components missing from existing libraries.
- Produce construction drawing sets to a firm’s standards.
- Transfer models to Revit with existing families retained.
- Revise its own output when a user identifies a problem, rather than requiring a rebuild by hand.
Projects can start from an empty site or from existing Revit, Rhino or SketchUp models, and firms can configure workflows against their own standards and knowledge bases. Snaptrude’s homepage separately advertises “Seamless Rhino & Revit export — Export .rvt files with no lost materials or broken parameters,” and says a model “complete with materials and parameters, and furniture can be downloaded or directly brought into Revit via Snaptrude Manager plugin.”
The AEC Magazine write-up does not carry pricing, but Snaptrude’s pricing page does, and it is the more revealing document of the two — because the agent is metered, not seated. Every tier lists an allowance of credits, and the “AI design agent” appears on the free tier: Free at $0 with one project, no exports and no BIM mode, and 75 credits a week on the annual view (the monthly view shows a one-time 300 credits instead); Pro at $50 a month billed yearly (or $60 monthly) with 500 credits a week, unlimited projects, BIM mode and exports; Organization at $100 per seat per month billed yearly (or $125 monthly) with 2,500 credits per seat per month plus teams, licence management and a BIM library; and Enterprise at a custom price with “credits tailored to your team.” The page does not say what one agent task costs in credits.
Snaptrude Inc.’s about page names Altaf Ganihar as “CEO & Founder” and describes “a growing team of ~100 people.” The most recent funding round this publication could find is a $14 million Series A announced on 9 November 2023, which the company’s own post says was led by Foundamental with Accel participating. Published totals for the company’s lifetime funding differ across trackers, so none is used here.
What AECBench Actually Measures
This is the part worth reading closely, because the methodology is set out in unusual detail for a vendor benchmark.
AECBench gives a model a design brief, has it design inside a BIM model, and scores the result stage by stage. The task categories are space planning, BIM detailing, drawings, services (plumbing, electrical, HVAC, fire) and value engineering. The example tasks Snaptrude lists are concrete rather than abstract: “Design a four-storey kindergarten,” “Lay out plumbing from the riser on an office floor,” “Six-unit Doppelhaus for maximum area efficiency,” “Furnish four rooms in an existing home” and “Retail presentation sheet.” Some tasks supply a starting site, massing or program.
Scoring runs in two tiers.
Core requirements, applied to every task, are five binary checks: no overlapping spaces, correct storey stacking, every space reachable, brief numbers met, and elements built as parametric components. These are checked by deterministic code reading the building model — not by a model judging a model.
Design aspects, applied where relevant, are the five categories above, each rated 1–5. These are scored by a review LLM. Snaptrude says practising architects wrote reviews and ratings that were used to build the review model’s criteria, that those reviews were split into a build set (60–70%) and a held-out test set (30–40%), and that agreement was measured using issue recall and precision, rating agreement and pairwise agreement. On the ceiling, the post’s own sentence is: “A subset of outputs was reviewed by two architects independently. Their agreement with each other set the ceiling against which the review LLM was judged.”
That last detail is the methodologically serious one. A human-ceiling baseline is how you stop an LLM grader’s agreement score from being read as an accuracy score. It is not unprecedented in construction AI benchmarks — Handoff’s takeoff benchmark publishes a score for independent human estimators, and Togal’s scoreboard carries a labelled human-in-the-loop reference mark — but in both of those cases the human number is a performance baseline on the task. Snaptrude’s is a baseline for the grader: it measures how much agreement two humans achieve on a subset of the same outputs, before an LLM reviewer is asked to match it.
The First Scores
Snaptrude’s post compares three systems: its own agent (which it says selects a model per task, Claude or GPT), Claude Code running Opus 5.5, and Codex running GPT-6 Astra. Scores are out of 5, and are presented in the order Snaptrude agent / Claude Code / Codex:
| Measure | Snaptrude | Claude Code | Codex |
|---|---|---|---|
| Space planning — brief fit | 3.7 | 3.5 | 3.5 |
| Space planning — adjacency and zoning | 4.0 | 3.6 | 3.2 |
| Space planning — proportions and efficiency | 3.3 | 3.2 | 2.8 |
| Space planning — site response | 3.0 | 3.6 | 3.4 |
| BIM detailing (walls, windows, doors, floors, roofs, edges) | 3.5 | 2.8 | 2.5 |
| Value engineering — area and proportions | 3.0 | 3.3 | 2.5 |
On the six measures published, Snaptrude’s agent posts the highest score on four and does not on two: site response, where Claude Code leads at 3.6 against Snaptrude’s 3.0, and value-engineering area and proportions, where Claude Code leads at 3.3 against Snaptrude’s 3.0. That tally is this publication’s arithmetic on Snaptrude’s published figures.
Two caveats are Snaptrude’s own, printed on the same page, and they matter for how the comparison should be read. The first is that the results come from Snaptrude’s internal runs. The second is that the two coding agents produced Three.js scenes rather than native building models — which is to say they were asked to do architectural work in a medium that is not a BIM model, and the gap on BIM detailing (3.5 against 2.8 and 2.5) is the measure where that difference would be expected to show up hardest.
Snaptrude says AECBench is not yet released, and that it plans to publish the tasks — briefs, setups, checks and applicable aspects — as open source, and to offer labs a way to run it on their own models inside Snaptrude. Until that happens, the scores above are not independently reproducible. Snaptrude says as much.
Four Benchmarks, and They Measure Different Things
Snaptrude’s stated reason for building AECBench is that “AEC professionals have no reliable way to judge whether an AI model can do the work they do,” and that existing evaluations “focus on general reasoning, coding and question answering.” Readers should have the wider record, because the field has become crowded quickly and the benchmarks do not measure the same thing.
- AEC-Bench, an academic benchmark with a nearly identical name, was posted to arXiv on 31 March 2026 by Harsh Mankodiya, Chase Gallik, Theodoros Galanos and Andriy Mulyar. It tests agentic systems on drawing understanding, cross-sheet reasoning and project-level coordination, and releases its dataset, harness and evaluation code under Apache 2.0. Its subject is reading documents, not producing designs.
- Construction’s Last Exam from Togal.AI and a rival benchmark from Handoff both score quantity takeoff. This newsroom covered them together on 19 October: two takeoff benchmarks landed eleven weeks apart and do not agree on how good the models are, with the top general-purpose model on Togal’s scoreboard at 34.0 out of 100 while Handoff’s puts frontier models in the 40s and 50s.
- AECBench scores something different from either: generated design, rather than document comprehension or quantity extraction. Reading a drawing has a right answer that can be checked against the drawing; counting a quantity has a right answer that can be checked against a bill. A floor plan does not, which is why the architect-written review criteria and the grader agreement ceiling are the load-bearing parts of Snaptrude’s design.
The analysis here is this publication’s own. A vendor publishing a benchmark on which its own product leads is the normal shape of this genre and does not by itself tell you anything; what distinguishes a useful one is whether the tasks, the checks and the grader are inspectable by someone else. Snaptrude has published the methodology and the first numbers and promised the tasks. Togal published a public scoreboard and a downloadable score file. AEC-Bench shipped the code. The direction of travel across all three is toward numbers a buyer can check, and a published number a competitor can dispute is a different kind of claim from a demo.
What to Watch
Two things will determine whether AECBench becomes a reference or a launch asset. The first is release: open-sourced tasks that a third party can run against a model Snaptrude does not control. The second is whether anyone other than Snaptrude publishes a score on it — a competitor submitting a run, or a lab reporting results, is the point at which a vendor benchmark becomes an industry one.
On the agent itself, the claim to test is LoD 400. The distance between a massing study and a coordinated model with structure and MEP at that level of development is where the work a firm actually bills for begins, and it is the part of the job a tool has to cross to be more than a concept aid. Snaptrude says the Agent crosses it in minutes.
The pricing page makes the second question sharper rather than answering it. A credit allowance is a consumption meter, and the figure that decides whether 500 credits a week is generous or tight is the credit cost of one agent task at LoD 400 — which this publication could not find published on the pricing page or in coverage of the launch. This is the same structural shift Motif made explicit when it priced its BIM authoring tool per AI user rather than per seat: once the expensive thing is inference rather than a human in a chair, the seat stops being the unit that makes sense to sell. The two have set the entry point in very different places. Snaptrude’s starts at $0 with a weekly credit allowance and the agent included; Motif’s published pricing starts at $150 per month per active user for Motif Design, lists no free plan, and its FAQ says “we don’t currently offer a trial for Motif Design” — though Motif Review, at $25 per seat per month, bundles 60 AI credits per seat per month, so the credit is becoming the unit on both sides. The commercial risk in giving the agent away at the bottom of the range is the same one every metered AI product carries, and it is a question about unit economics rather than about architecture.
