AI construction takeoff leaderboard

Updated November 17, 2026

  • OpenAI
  • Anthropic
  • Google
  • Kimi
  • DeepSeek
  • Z.ai
  • Qwen

Takeoff rating

Each model at its highest-rated thinking level · Higher is better

How the takeoff rating works

The takeoff rating is out of 100: accuracy contributes 50 points, cost 25 and speed 25.

For each drawing, we estimate how long the takeoff would take by hand and use an hourly labor rate to estimate its cost. Each thinking level is rated separately against those estimates.

Scoring details
  • Accuracy: 85% earns 0 points; 100% earns all 50. Below 85%, the overall rating halves for every 5 percentage points lost.
  • Cost: no points for matching or exceeding the estimated manual cost; full points at 1/100 of that cost.
  • Speed: no points for matching or exceeding the estimated manual time; full points at 1/100 of that time.

We estimate an average of 30 minutes of manual work per test across the 4 selected tests, costing 1,916 credits ($19.16) at an estimator wage of $38 per hour.

Time matters because when you use AI for a takeoff, you work alongside the model: reviewing its work, requesting corrections and waiting for each response. For unattended takeoffs, you may prefer to prioritize accuracy and cost.

The tags mark the leaders among models scoring 85% or more: Most accurate for the highest score, Best value for the most accuracy per credit, and Best speed for the most accuracy per minute.

How accurate

Score at each model’s most accurate level · Higher is better

How fast

Minutes at each model’s most accurate level · Lower is better

How cheap

Credits at each model’s most accurate level · Lower is better

Score floor for the takeoff rating and value charts:3 models score less, so the value charts leave them out.

Value for money

Score per credit at each model’s most accurate level · Higher is better

Value for time

Score per minute at each model’s most accurate level · Higher is better

Accuracy vs

Score against credits per takeoff for every level each model ran. Cost runs on a log scale, and score on one that spreads out high scores. A line joins one model's levels. Up and to the left is better.

Does more thinking help?

The effect of more thinking varies by model. Sometimes it buys a small accuracy gain at a much higher cost in time and money. Other models become faster or cheaper at higher thinking levels. Compare the results for each model rather than assuming a lower setting is more efficient.

About this benchmark

We test AI models on construction PDFs: measuring floor areas and counting doors, windows and electrical devices.

We compare each result with an answer key and combine accuracy, cost and time into a takeoff rating out of 100.

Claude Fable 5.2 leads on accuracy. Terra 6.1 has the highest overall rating when cost and speed are included.

How we test

Each test includes a real construction drawing, a takeoff request and an answer key.

We test models at different thinking levels and record:

  • Accuracy: how closely the markups match the answer key. Incorrect categories, missing items and double counting lose points.
  • Cost: the Deta credit charge for each run. One credit is $0.01.
  • Time: from sending the request to the finished markups.

Valid alternatives are accepted: a stair area can be counted on either floor it connects, but only once.

Each test has equal weight. Repeated runs are averaged within each test; results from older test versions are excluded.

These results cover 4 drawing tests using Deta’s takeoff tools. Performance can vary with other drawings, prompts and tools. Manual times are estimates, not timed human trials.

Read the benchmark findings and see example drawings

Choosing a model

Use the overall rating as a starting point. If accuracy matters most, compare the accuracy scores directly. For unattended work, cost may matter more than speed. Check the results for the type of takeoff you need, and review the finished markups.

Deta’s model picker shows the takeoff rating for each thinking level, alongside accuracy, cost and speed, so you can use these results when choosing a model.

Models

Rows show each model’s most accurate level, prioritizing levels with more completed tests. Expand a row to compare thinking levels.

ModelLevelScoreTestsRunsDurationCreditsBy test
Claude Fable 5.2Most accurate
Low99.7%4/4438m 45s1458Area_Complex_1 99.5% · Area_Simple_1 99.3% · Electrical_Simple_1 100% · Opening_Simple_1 100%
Astra 6
High99.6%4/449m 25s554Electrical_Simple_1 100% · Opening_Simple_1 100% · Area_Complex_1 98.4% · Area_Simple_1 99.9%
Sol 6.1Best speed
High99.3%4/447m 21s77Electrical_Simple_1 99.6% · Area_Simple_1 99.7% · Opening_Simple_1 100% · Area_Complex_1 97.9%
Terra 6.1
Maximum99.3%4/4417m 25s103Area_Complex_1 97.2% · Opening_Simple_1 100% · Electrical_Simple_1 100% · Area_Simple_1 99.8%
Gemini 4 Carbon
High99.3%4/4452m 33s494Area_Complex_1 97.2% · Electrical_Simple_1 100% · Area_Simple_1 99.8% · Opening_Simple_1 100%
Gemini 4 Argon
High98.6%4/4437m 23s393Opening_Simple_1 100% · Electrical_Simple_1 100% · Area_Simple_1 98.9% · Area_Complex_1 95.6%
Claude Opus 5.5
High95.3%4/4419m 22s352Opening_Simple_1 100% · Area_Simple_1 95.9% · Electrical_Simple_1 92.2% · Area_Complex_1 93%
Claude Sonnet 5.5
Medium94.4%4/4444m 51s448Opening_Simple_1 100% · Area_Complex_1 88.7% · Area_Simple_1 96.8% · Electrical_Simple_1 92%
Kimi K3.1 Pro
High92%4/4411m 51s132Area_Complex_1 78.5% · Electrical_Simple_1 96.6% · Area_Simple_1 92.8% · Opening_Simple_1 100%
Qwen4 Max
High89.7%4/4435m 35s156Area_Complex_1 88.6% · Electrical_Simple_1 99.6% · Opening_Simple_1 75% · Area_Simple_1 95.6%
Luna 6Best value
Maximum88.6%4/4426m 38s22Electrical_Simple_1 97.3% · Area_Simple_1 95.7% · Opening_Simple_1 75% · Area_Complex_1 86.2%
Gemini 3.8 Flash
Low87%4/4417m 10s237Area_Simple_1 98.7% · Area_Complex_1 64.6% · Opening_Simple_1 95% · Electrical_Simple_1 89.7%
GLM 5.3 FlashX
Low83.7%4/4416m 54s99Area_Simple_1 86.4% · Electrical_Simple_1 81.1% · Opening_Simple_1 92.7% · Area_Complex_1 74.4%
DeepSeek V4.1 Flash
High80.1%4/4413m 08s15Area_Complex_1 67.9% · Opening_Simple_1 68.3% · Area_Simple_1 91.6% · Electrical_Simple_1 92.6%
Claude Haiku 5.5
Low70.8%4/447m 59s9Area_Complex_1 61.5% · Area_Simple_1 86.9% · Electrical_Simple_1 73.5% · Opening_Simple_1 61.1%