Can AI Read a Blueprint Yet? What Blueprint-Bench 2 Shows

If you handed a floor plan to ChatGPT or Claude a year ago and asked for a takeoff, you probably closed the tab five minutes later and went back to Bluebeam. It was fine for summarizing a spec section or cleaning up an RFI, but it couldn’t read a drawing well enough to trust on a bid.

That’s why an entire category of standalone AI takeoff software popped up. General-purpose models couldn’t trace a wall or read a scale, so vendors like Togal.AI built custom computer vision tools to do that one job. (We looked at several of them in our guide to AI preconstruction software.)

Over the last couple of months, the ground shifted under that whole category.

Today, you can drop an architectural set into Gemini 4 Argon, Claude Opus 5.5, or GPT-6 Astra and get a real takeoff back in minutes. And it’s not just counting door tags or reading schedule tables. These models read the scale off the sheet and are genuinely good at linear and area measurements too: linear feet of walls broken out by partition type, square footage of flooring or ACT ceilings, slab areas, and fixture counts, all cross-referenced against the schedules and tied back to the sheet each number came from. We’ve been throwing real plan sets at them at Buildr, and the results are genuinely good.

The Proof Is in the Curve

There isn’t a public benchmark specifically for construction takeoff yet, so the best way to see what changed is Blueprint-Bench 2 from Andon Labs.

Instead of giving the AI a finished drawing, Blueprint-Bench 2 gives it a handful of photos taken inside an apartment and asks it to draw a scaled floor plan from scratch. That’s actually a harder spatial and measurement problem than takeoff. On a blueprint, the architect already did the hard part: the walls are drawn to scale, the dimensions are called out, and the rooms and doors are tagged.

Best Blueprint-Bench 2 score, by model release date

The green line is the best score released so far. Gray dots are every other model tested.

Source: Andon Labs, Blueprint-Bench 2, retrieved October 12, 2026. Scores are normalized so random guessing is 0 and a perfect plan is 1. The human baseline was measured on 12 of the 50 apartments.

Look at what happened on that chart this year. In February, the best model scored 0.27. By the end of September, Google’s Gemini 4 Argon pushed the top score to 0.54—almost touching the 0.59 human baseline. And notice how it happened: the line didn’t creep up slowly. Almost the entire jump happened in a few weeks in September as GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon shipped one after another.

If you tried AI on a plan set back in the spring and wrote it off because it botched the measurements, you were testing a completely different generation of models.

Stop Buying Point Solutions for a Moving Target

When spatial reasoning and measurement are improving that fast at the model layer, buying a dedicated takeoff tool is a bad bet.

Standalone takeoff products take months or years to train and package their own vision models. By the time you buy one, roll it out, and train your estimators on a new interface, the frontier models from Google, Anthropic, and OpenAI have already leapfrogged it.

You’re much better off working as close to the frontier models as you can:

In Gemini, Claude, or ChatGPT directly, especially if your estimating tool or spreadsheet supports MCP so the model can drop quantities straight into your estimate instead of making you copy and paste tables out of a chat window.
In Kit, our AI agent inside Buildr, which runs on the newest frontier models and drafts takeoff quantities (counts, linear feet, and square footages) directly as line items in your budget, right alongside your historical costs from past jobs.

Either way, when the next model drops four to six weeks from now, your takeoff workflow gets better that afternoon. You don’t have to wait on a vendor’s product roadmap or migrate to another tool.

Sending ITBs by hand.
Getting bids by email.
Leveling in Excel.

There's only one AI-powered platform that handles all of it end-to-end. You're looking at it.

Start Trial Contact

Why Waiting Costs More Than It Used To

Every time we show this to a preconstruction team, someone asks if they should just wait six months until the line on that chart crosses the human baseline.

The problem with waiting is that using these models well is a skill. The estimators experimenting right now are learning how to break up a 150-sheet set, how to check linear and area takeoffs quickly, which sheets still need a closer human eye, and how to prompt the model so the output matches how their company actually builds a budget. When the next model ships, those teams plug it into a workflow that’s already running.

The easiest way to start isn’t a pilot program or a software rollout. Grab a plan set your team already bid by hand last month, upload a few sheets to Gemini, Claude, ChatGPT, or Kit, and compare the counts, wall lengths, and floor areas against your own takeoff. You’ll know in an afternoon where it saves your estimators two days of tracing—and where they can spend that time on scope gaps, sub coverage, and pricing instead.

If you want to see Kit run a takeoff and build a draft budget inside Buildr, book a demo and bring a set you’ve already bid.

FAQ

Can ChatGPT, Claude, or Gemini do a quantity takeoff?

Yes. Frontier models like Gemini 4 Argon, Claude Opus 5.5, and GPT-6 Astra can read a sheet’s title block and scale, count doors and fixtures, measure linear feet of walls and square footage of rooms or finishes, and pull types from the schedules. Just make sure you use a business or enterprise plan that doesn’t train on your files.

How do the numbers get from a chat app into my estimate?

Through MCP (Model Context Protocol), an open standard that lets AI tools connect directly to other software. If your estimating software, spreadsheet, or ERP supports MCP, the chat app can write quantities straight into your estimate as line items.

Can Kit do takeoff inside Buildr?

Yes. Kit is the AI agent built into Buildr. When you upload a set of plans, Kit reads the sheets and drafts the counts, linear measurements, and areas as line items directly in your Buildr budget, right next to your historical project costs.

Do I still need dedicated AI takeoff software?

Probably not. General-purpose frontier models now handle the counting, linear, and area measurements those standalone tools were built for, and a better model ships every four to six weeks. You’re better off working with the models directly in Gemini, Claude, ChatGPT, or Kit.