Togal.AI has published a benchmark that tests general-purpose AI models on construction quantity takeoff, and the best one it ran got 34% right. The Miami software company sells its own machine-vision takeoff tool, so the scoreboard comes with an obvious interest attached.
Togal calls the test Construction’s Last Exam. Construction Briefing reported the results Oct. 13. The benchmark has 33 tasks across three families: 12 area tasks that trace rooms, finishes and paving, 12 counting tasks that find doors, fixtures and devices, and nine linear tasks that pull wall centerlines and outlines.
How the AI takeoff benchmark works
Each model gets raw PDF drawings. It can use its built-in vision and code execution but no third-party tools and no agentic workflows. Togal scores the output against ground-truth geometry using standard overlap and detection measures (intersection over union, F1 and mean average precision), then averages the three families into one number.
Togal’s published score file lists two dozen models. The top of the table, per Togal:
- OpenAI’s GPT-6 Sol: 34.0%
- Google’s Gemini 4 Argon, dated Oct. 13: 25.7%
- OpenAI’s GPT-6 Luna: 21.6%
- Gemini 3.8 Flash and GPT-6 Astra: 19.2% each
- Anthropic’s Claude Opus 5.5: 16.2%
The bottom sits under 2%. Togal’s own entry scores 97.7%, but it isn’t a model run. The site labels it a reference mark with a human in the loop for up to 30 seconds a task.
Read the construction AI scores with the source in mind
Three caveats matter. Togal built the tasks, and its internal estimators validated the ground truth. The rules strip out the tools and agent setups that most AI products wrap around a model, which is a fair test of raw drawing-reading but not of what a vendor actually ships. And the human-assisted 97.7% isn’t comparable to an unassisted model score, even on the same chart.
To Togal’s credit, it published the score file, task instances and a GitHub repository, so others can rerun it. That’s more than most vendor claims offer. It’s also labeled a geometry-only release, with a fuller suite that adds reasoning and end-to-end tasks still in preview.
What the drawing-reading gap means for estimators
The trend line is the real news. In Togal’s own data, models released in 2025 mostly scored between 2% and 4%. The best September 2026 release hit 34%. That’s a big jump in about a year, even if it’s still a failing grade for any estimator who has to sign a bid.
It lines up with what firms report in practice. In BRG’s construction AI survey, more firms said they added people because of AI (18%) than cut staff (11%). Tools sold on drafting speed, like Blueprints AI’s permit-drawing product, still need someone checking the sheets.
For preconstruction teams, the useful step is simple. Run the public tasks, or a few of your own sheets, through whatever general model your staff already use, and compare it to a manual count before anyone trusts it on a bid. If the next round of frontier models clears 50% on Togal’s test, the case for specialized takeoff software gets harder to make on accuracy alone.