Togal.AI's takeoff benchmark scores the best general AI model at 34.0 out of 100 on 33 drawing tasks

What it does

Each of the 33 tasks gives a model one drawing and one instruction

Each task gives a model a drawing and one instruction, and the model has to return geometry. Togal sorts the 33 tasks into three families, and the overall score is the mean of the three family scores.

Evidence so far

Togal wrote the 33 tasks, ran the models and scored its own platform at 97.7

The evidence is Togal's own benchmark site and its press release. Togal wrote the tasks, its estimators validated the answer key, and it ran the models and published the scores, its own platform included. Digital Journal reported that the results have not yet been checked through peer review or by a third party. The rule for grading a traced polygon or a located symbol against the answer key is not described on the overview page.

The table supports these readings.

1. The spread among general models is wide. GPT-6 Sol at 34.0 led GPT-6 Luna, in second among general models, by 12.4 points. The rest of the first page sits between 12.3 and 21.6.
2. Cost does not line up with score. Togal scales each provider's published API rates, the per-use charges for sending work to a model through its programming interface, so that one full run of GPT-6 Astra costs about $14.0. On that scale GPT-6 Sol cost about $3.5 and scored 34.0, while GPT-6 Astra scored 19.2.

How to measure it on your projects

Test a tool on the 10 to 15 costliest line items from two or three recent bids

Anyway, the exam runs on drawings Togal chose, and your own sheets are the better test for your trades. If you pilot any AI takeoff tool, whether a general model or a product such as Quotr or STACK IQ, track quantity variance by line item. It is the difference between the tool's quantity and the estimator's checked quantity, divided by the estimator's quantity, for each line such as LF of drywall partition, SF of flooring, or count of receptacles.

Capture the baseline before the pilot:

1. Pick two or three recent bids where the takeoff was checked and the job went ahead, so field quantities can confirm the estimate.
2. Record the estimator's hours on takeoff for each bid and the quantity for 10 to 15 line items that carry the most cost.
3. Run the same sheets through the tool and compute the variance for each line, plus the estimator's time to review and correct the tool's output.

Split the results the way Togal does, into linear, area, and count items, since the models in the exam scored differently on each. If a tool counts fixtures well and misses wall centerlines, you can use it on count items and keep linear takeoff manual. A first step on one job is to take a single floor plan from a recent bid and have the estimator compare the tool's receptacle count and drywall LF against the numbers that went into the bid.

Availability and cost

One model's run through all 33 tasks cost about $0.2 to $14.0 in API charges

The benchmark site is public at cle.togal.ai, with a page listing all 33 task prompts. Togal says the drawing PDFs, its own baseline output, and the general models' run files are in a public Google Drive folder. On the first page of the leaderboard, one model's run through all 33 tasks ranged from about $0.2 to about $14.0 in API charges. The GitHub repository linked from the site returned a not-found page when this post was written on October 8. Togal's overview page marks this as a geometry-only release and links a preview of a fuller suite covering reasoning and end-to-end tasks.