Benchmarks

How well can AI read construction drawings?

169 questions an estimator wrote against 9 real sets. We gave them to 30 frontier models through the raw API — no tools, no help, just the model and the drawing — and then we ran them again inside a harness (Cowork, ChatGPT Work, Claude Code, Cursor, Codex): 14 model-and-harness pairs, with the drawings in a folder they could open.

169

Questions

Real drawing sets

Raw models

Harness + model

82.1%

Best model alone

92.5%

Best in a harness

The same questions, inside a harness

Agents working the drawings out of a folder, the way you would. Click a row for its category breakdown.

#	Harness	Model	Accuracy	vs raw	Details
1	Cowork	Claude Fable 5.1	92.5%	+22.4	›
Cowork · Claude Fable 5.1 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 70.2%. Reading the sheet 100.0% 29 questions Schedules 92.3% 13 questions Counting 91.2% 34 questions Element detail 95.2% 21 questions Scale and measurement 72.7% 11 questions Cross-referencing 92.3% 13 questions Scope 100.0% 6 questions Sheet structure 85.7% 7 questions Knowing when to stop 33.3% 3 questions
2	Cowork	Claude Opus 5	91.8%	+23.9	›
Cowork · Claude Opus 5 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 67.9%. Reading the sheet 100.0% 29 questions Schedules 100.0% 13 questions Counting 88.2% 34 questions Element detail 95.2% 21 questions Scale and measurement 81.8% 11 questions Cross-referencing 92.3% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 33.3% 3 questions
3	ChatGPT Work	GPT 5.6 Sol	90.3%	+11.2	›
ChatGPT Work · GPT 5.6 Sol by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 79.1%. Reading the sheet 96.5% 29 questions Schedules 100.0% 13 questions Counting 85.3% 34 questions Element detail 100.0% 21 questions Scale and measurement 90.9% 11 questions Cross-referencing 76.9% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 33.3% 3 questions
4	Claude Code	Claude Opus 5	88.1%	+20.2	›
Claude Code · Claude Opus 5 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 67.9%. Reading the sheet 93.1% 29 questions Schedules 100.0% 13 questions Counting 82.3% 34 questions Element detail 90.5% 21 questions Scale and measurement 63.6% 11 questions Cross-referencing 84.6% 13 questions Scope 100.0% 6 questions Sheet structure 100.0% 7 questions Knowing when to stop 0.0% 3 questions
5	ChatGPT Work	GPT 6 Astra	86.6%	+4.5	›
ChatGPT Work · GPT 6 Astra by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 82.1%. Reading the sheet 86.2% 29 questions Schedules 100.0% 13 questions Counting 79.4% 34 questions Element detail 95.2% 21 questions Scale and measurement 63.6% 11 questions Cross-referencing 92.3% 13 questions Scope 83.3% 6 questions Sheet structure 100.0% 7 questions Knowing when to stop 33.3% 3 questions
6	Cursor	Kimi K3	83.6%	—	›
Cursor · Kimi K3 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. Reading the sheet 96.5% 29 questions Schedules 100.0% 13 questions Counting 73.5% 34 questions Element detail 85.7% 21 questions Scale and measurement 54.5% 11 questions Cross-referencing 92.3% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 33.3% 3 questions
7	Cowork	Claude Sonnet 5	82.8%	+22.4	›
Cowork · Claude Sonnet 5 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 60.5%. Reading the sheet 96.5% 29 questions Schedules 100.0% 13 questions Counting 70.6% 34 questions Element detail 90.5% 21 questions Scale and measurement 45.5% 11 questions Cross-referencing 84.6% 13 questions Scope 83.3% 6 questions Sheet structure 85.7% 7 questions Knowing when to stop 100.0% 3 questions
8=	Cursor	Claude Opus 5	82.1%	+14.2	›
Cursor · Claude Opus 5 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 67.9%. Reading the sheet 96.5% 29 questions Schedules 84.6% 13 questions Counting 76.5% 34 questions Element detail 81.0% 21 questions Scale and measurement 54.5% 11 questions Cross-referencing 92.3% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 33.3% 3 questions
9=	Cursor	GLM 5.2	82.1%	—	›
Cursor · GLM 5.2 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. Reading the sheet 96.5% 29 questions Schedules 92.3% 13 questions Counting 79.4% 34 questions Element detail 90.5% 21 questions Scale and measurement 45.5% 11 questions Cross-referencing 69.2% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 33.3% 3 questions
10	Codex	GPT 5.6 Sol	80.6%	+1.5	›
Codex · GPT 5.6 Sol by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 79.1%. Reading the sheet 96.5% 29 questions Schedules 92.3% 13 questions Counting 67.7% 34 questions Element detail 95.2% 21 questions Scale and measurement 54.5% 11 questions Cross-referencing 61.5% 13 questions Scope 83.3% 6 questions Sheet structure 85.7% 7 questions Knowing when to stop 0.0% 3 questions
11=	Cursor	Grok 4.6	79.8%	+30.6	›
Cursor · Grok 4.6 by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 49.3%. Reading the sheet 93.1% 29 questions Schedules 100.0% 13 questions Counting 70.6% 34 questions Element detail 81.0% 21 questions Scale and measurement 63.6% 11 questions Cross-referencing 69.2% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 0.0% 3 questions
12=	Cursor	GPT 5.6 Sol	79.8%	+0.8	›
Cursor · GPT 5.6 Sol by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 79.1%. Reading the sheet 93.1% 29 questions Schedules 76.9% 13 questions Counting 70.6% 34 questions Element detail 85.7% 21 questions Scale and measurement 63.6% 11 questions Cross-referencing 76.9% 13 questions Scope 83.3% 6 questions Sheet structure 85.7% 7 questions Knowing when to stop 33.3% 3 questions
13	Cursor	Gemini 3.7 Flash	76.9%	+6.7	›
Cursor · Gemini 3.7 Flash by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 70.2%. Reading the sheet 93.1% 29 questions Schedules 100.0% 13 questions Counting 55.9% 34 questions Element detail 90.5% 21 questions Scale and measurement 45.5% 11 questions Cross-referencing 61.5% 13 questions Scope 100.0% 6 questions Sheet structure 85.7% 7 questions Knowing when to stop 33.3% 3 questions
14	ChatGPT Work	GPT 5.6 Luna	69.4%	+1.5	›
ChatGPT Work · GPT 5.6 Luna by category All 169 questions in one sitting. The score above covers the 134 of those the drawings can actually answer. The same model straight off the API scored 67.9%. Reading the sheet 96.5% 29 questions Schedules 69.2% 13 questions Counting 47.1% 34 questions Element detail 90.5% 21 questions Scale and measurement 45.5% 11 questions Cross-referencing 46.2% 13 questions Scope 83.3% 6 questions Sheet structure 71.4% 7 questions Knowing when to stop 0.0% 3 questions

One session per pairing, run once. Repeat runs of this benchmark move by 1.1 to 5.4 points on their own, so read anything closer than that as a tie rather than an order. Agents worked however they liked, and some built themselves crops, OCR dumps and index files along the way.

And the raw models on their own

No tools, no folder, no second look. Accuracy is the share of answerable questions it got right. Click a row for its category breakdown.

#	Model	Lab	Accuracy	Cost	Details
1	GPT 6 Astra openai/gpt-6-astra	OpenAI	82.1%	$24.11	›
GPT 6 Astra by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 6.7% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 92.3% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 73.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 95.2% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 45.5% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 84.6% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
2	GPT 5.6 Sol openai/gpt-5.6-sol	OpenAI	79.1%	$13.54	›
GPT 5.6 Sol by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 1.5% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 100.0% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 64.7% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 36.4% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 76.9% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
3	GPT 5.6 Terra openai/gpt-5.6-terra	OpenAI	76.1%	$5.56	›
GPT 5.6 Terra by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 3.7% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 86.2% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 100.0% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 70.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 81.0% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 84.6% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
4	Gemini 3.8 Flash google/gemini-3.8-flash	Google	71.6%	$1.18	›
Gemini 3.8 Flash by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 14.9% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 96.5% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 100.0% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 47.1% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 81.0% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 69.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 33.3% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
5=	Claude Fable 5.1 anthropic/claude-fable-5.1	Anthropic	70.2%	$15.05	›
Claude Fable 5.1 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 14.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 76.9% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 35.3% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 36.4% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 84.6% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 100.0% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 66.7% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
5=	Gemini 3.7 Flash google/gemini-3.7-flash	Google	70.2%	$0.33	›
Gemini 3.7 Flash by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 11.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 100.0% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 41.2% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 53.8% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 100.0% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
7	GPT 5.5 openai/gpt-5.5	OpenAI	69.4%	$19.90	›
GPT 5.5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 6.0% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 92.3% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 42.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 27.3% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 76.9% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 35.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
8=	Claude Opus 5 anthropic/claude-opus-5	Anthropic	67.9%	$7.45	›
Claude Opus 5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 20.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 76.9% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 35.3% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 36.4% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 61.5% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
8=	GPT 5.6 Luna openai/gpt-5.6-luna	OpenAI	67.9%	$0.41	›
GPT 5.6 Luna by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 5.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 89.7% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 92.3% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 50.0% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 81.0% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 69.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 42.9% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 66.7% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
10=	Claude Fable 5 anthropic/claude-fable-5	Anthropic	65.7%	$14.71	›
Claude Fable 5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 13.4% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 93.1% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 84.6% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 32.4% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 53.8% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 100.0% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 66.7% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
10=	Claude Opus 4.8 anthropic/claude-opus-4.8	Anthropic	65.7%	$7.06	›
Claude Opus 4.8 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 27.6% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 89.7% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 76.9% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 26.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 90.5% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 27.3% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 69.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
12	Gemini 3.6 Flash google/gemini-3.6-flash	Google	63.4%	$3.38	›
Gemini 3.6 Flash by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 14.5% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 84.5% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 92.3% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 27.9% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 83.3% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 13.6% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 69.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
13	Gemini 3.1 Pro Preview google/gemini-3.1-pro-preview	Google	62.7%	$2.46	›
Gemini 3.1 Pro Preview by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 18.7% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 86.2% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 92.3% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 29.4% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 81.0% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 61.5% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 71.4% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
14	Claude Sonnet 5 anthropic/claude-sonnet-5	Anthropic	60.5%	$2.94	›
Claude Sonnet 5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 26.9% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 89.7% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 69.2% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 26.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 46.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
15	Claude Opus 4.7 anthropic/claude-opus-4.7	Anthropic	57.5%	$6.48	›
Claude Opus 4.7 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 23.9% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 86.2% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 76.9% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 20.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 85.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 61.5% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 66.7% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
16=	Gemini 3.1 Flash Lite google/gemini-3.1-flash-lite	Google	56.7%	$0.08	›
Gemini 3.1 Flash Lite by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 14.9% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 82.8% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 69.2% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 26.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 76.2% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 61.5% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 42.9% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
16=	GPT 5.4 Mini openai/gpt-5.4-mini	OpenAI	56.7%	$2.28	›
GPT 5.4 Mini by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 11.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 75.9% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 76.9% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 35.3% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 71.4% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 53.8% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
18=	Claude Sonnet 4.5 anthropic/claude-sonnet-4.5	Anthropic	51.5%	$4.14	›
Claude Sonnet 4.5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 28.4% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 86.2% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 53.8% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 20.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 71.4% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 38.5% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 66.7% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
18=	GPT 4.1 openai/gpt-4.1	OpenAI	51.5%	$1.61	›
GPT 4.1 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 11.9% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 86.2% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 53.8% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 23.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 81.0% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 38.5% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 100.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 14.3% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
20=	Gemini 2.5 Pro google/gemini-2.5-pro	Google	49.3%	$2.29	›
Gemini 2.5 Pro by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 14.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 79.3% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 76.9% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 23.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 57.1% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 30.8% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 66.7% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
20=	Grok 4.6 x-ai/grok-4.6	xAI	49.3%	$10.71	›
Grok 4.6 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 24.6% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 65.5% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 61.5% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 32.4% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 38.1% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 18.2% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 53.8% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 85.7% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 33.3% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
22=	Claude Haiku 4.5 anthropic/claude-haiku-4.5	Anthropic	46.3%	$1.57	›
Claude Haiku 4.5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 34.3% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 82.8% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 46.2% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 14.7% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 76.2% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 30.8% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 28.6% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
22=	Gemini 2.5 Flash google/gemini-2.5-flash	Google	46.3%	$0.36	›
Gemini 2.5 Flash by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 14.9% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 82.8% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 61.5% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 20.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 52.4% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 46.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 50.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 28.6% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
24	Qwen3 VL 235b A22b Instruct qwen/qwen3-vl-235b-a22b-instruct	Alibaba	44.8%	$1.18	›
Qwen3 VL 235b A22b Instruct by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 36.6% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 82.8% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 46.2% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 23.5% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 66.7% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 15.4% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 14.3% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
25	GPT 5 Nano openai/gpt-5-nano	OpenAI	41.8%	$0.19	›
GPT 5 Nano by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 38.8% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 75.9% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 46.2% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 17.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 57.1% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 23.1% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 83.3% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 14.3% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
26	Grok 4.5 x-ai/grok-4.5	xAI	40.3%	$2.12	›
Grok 4.5 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 25.4% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 58.6% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 61.5% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 14.7% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 42.9% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 9.1% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 46.2% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 66.7% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 57.1% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
27	GPT 4o openai/gpt-4o	OpenAI	38.8%	$1.79	›
GPT 4o by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 33.6% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 72.4% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 38.5% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 11.8% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 71.4% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 15.4% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 66.7% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 14.3% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
28	Gemini 2.5 Flash Lite google/gemini-2.5-flash-lite	Google	37.3%	$0.08	›
Gemini 2.5 Flash Lite by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 26.1% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 65.5% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 46.2% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 17.6% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 47.6% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 23.1% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 50.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 42.9% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 66.7% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
29=	Mistral Large 2512 mistralai/mistral-large-2512	Mistral	9.0%	$0.13	›
Mistral Large 2512 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 67.2% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 13.8% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 0.0% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 8.8% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 9.5% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 15.4% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 16.7% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 0.0% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.
29=	Mistral Medium 3.1 mistralai/mistral-medium-3.1	Mistral	9.0%	$0.12	›
Mistral Medium 3.1 by category Answered all 169 questions. The score above covers the 134 of those the drawings can actually answer. It declined 73.5% of them, which is the number worth watching: a model that will not commit is as much of a problem on site as one that guesses. Reading the sheet 13.8% 29 questions Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it. Schedules 7.7% 13 questions Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet. Counting 8.8% 34 questions Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first. Element detail 9.5% 21 questions The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is. Scale and measurement 0.0% 11 questions Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself. Cross-referencing 7.7% 13 questions Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer. Scope 0.0% 6 questions What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s. Sheet structure 14.3% 7 questions Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at. Knowing when to stop 100.0% 3 questions Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.

Cost is one full run of all 169 questions, model usage only. Scored on the native PDF lane. Last updated September 2026.

The questions

Written by an estimator

169 questions across architectural, structural, electrical, mechanical and plumbing sets. Every answer was worked out and checked by a practising estimator before any model saw the drawings.

The test

Same drawings, same questions

Every model and every agent gets the identical question and the identical scoring. No tuning toward any vendor, and nothing pre-digested. What differs between the two tables is the sitting: a raw model is handed one page, an agent is handed the folder and left to work.

The harnesses

Whatever the harness would really do

Each agent got one session and was left alone in it. Some worked straight off the PDFs; some built themselves image crops, OCR text dumps and index files first. That is part of what a harness is, so it is left in — which does mean two harnesses on the same model are not a clean like-for-like.

The scoring

Right or wrong, nothing in between

An answer is correct or it is not. Half a right answer to a three-part question is a wrong answer to it. Three of the questions cannot be answered from the drawings at all, and saying so is the correct response.

The categories

What we asked them to do

The questions split into 9 kinds of work. They are not equally hard, and the gap between them is the useful part: a model can be excellent at reading a schedule and still be unable to count.

Reading the sheet

32 questions

Finding a fact that is written down somewhere. Title block, general notes, the legend. The drawing says it outright, the model just has to go and find it.

Schedules

18 questions

Pulling one value out of a table. Door schedules, panel schedules, equipment lists. Easy to eyeball, easy to get wrong when the table runs across half a sheet.

Counting

41 questions

Counting tagged items on a plan. How many of this fitting, how many of that outlet. This is the core takeoff skill and the one most people want to hand over first.

Element detail

25 questions

The size, type or material of a named element. What that beam is, what that pipe is made from, what size that cable run is.

Scale and measurement

22 questions

Distances, areas and volumes that need scale reasoning. The number is not printed anywhere, so the model has to work it off the drawing itself.

Cross-referencing

14 questions

Answers that need more than one sheet, or tracing a single line diagram through. Nothing on its own page gives you the answer.

Scope

7 questions

What is in and what is out. New versus existing, who supplies it, whether it sits in this package or someone else’s.

Sheet structure

7 questions

Reading the drawing as a document rather than a picture. How many details are on the sheet, how many grid lines, what revision it is at.

Knowing when to stop

3 questions

Questions the drawings genuinely do not answer. The right response is to say so. A model that invents a number here is the one that costs you money.

We re-run this on every major release

When a new model ships it goes through the same 169 questions. Join the community and you get the results as they come in, plus the workflows we build off the back of them.

Get the results first →

How well can AI read construction drawings?

The same questions, inside a harness

Cowork · Claude Fable 5.1 by category

Cowork · Claude Opus 5 by category

ChatGPT Work · GPT 5.6 Sol by category

Claude Code · Claude Opus 5 by category

ChatGPT Work · GPT 6 Astra by category

Cursor · Kimi K3 by category

Cowork · Claude Sonnet 5 by category

Cursor · Claude Opus 5 by category

Cursor · GLM 5.2 by category

Codex · GPT 5.6 Sol by category

Cursor · Grok 4.6 by category

Cursor · GPT 5.6 Sol by category

Cursor · Gemini 3.7 Flash by category

ChatGPT Work · GPT 5.6 Luna by category

And the raw models on their own

GPT 6 Astra by category

GPT 5.6 Sol by category

GPT 5.6 Terra by category

Gemini 3.8 Flash by category

Claude Fable 5.1 by category

Gemini 3.7 Flash by category

GPT 5.5 by category

Claude Opus 5 by category

GPT 5.6 Luna by category

Claude Fable 5 by category

Claude Opus 4.8 by category

Gemini 3.6 Flash by category

Gemini 3.1 Pro Preview by category

Claude Sonnet 5 by category

Claude Opus 4.7 by category

Gemini 3.1 Flash Lite by category

GPT 5.4 Mini by category

Claude Sonnet 4.5 by category

GPT 4.1 by category

Gemini 2.5 Pro by category

Grok 4.6 by category

Claude Haiku 4.5 by category

Gemini 2.5 Flash by category

Qwen3 VL 235b A22b Instruct by category

GPT 5 Nano by category

Grok 4.5 by category

GPT 4o by category

Gemini 2.5 Flash Lite by category

Mistral Large 2512 by category

Mistral Medium 3.1 by category