model

Can ChatGPT Write a Construction Estimate? 12 Runs of One Bathroom Against Clamp (2026)

The same 5x8 bathroom remodel run through free and paid ChatGPT 12 times and checked line by line against Clamp: the totals, the math, the rates, the scope it dropped, and the prompt that works.

By Clamp ResearchReviewed by Emmanuel ForgePublished Nov 1, 2026✓ Reviewed Nov 2026

Direct answer: Yes, as a first draft, but not as a price. We gave ChatGPT the same 5x8 bathroom remodel 12 times. Free ChatGPT came back between $9,333 and $14,527, 56% apart, with math errors in two of four answers. The paid models landed −14.1% to −5.3% against Clamp's estimate of $13,858, under it every time, on rates ChatGPT made up and a scope that changed from run to run and left out work Clamp carries. Asked to price only from sources it could name, it listed 89 to 102 scope items and priced 2 or 3 of them.

The assumptions

This model assumes

One 5x8 bathroom, same layout, described the same way every run, no location
Totals before markup, sales tax and permits
Benchmark: Clamp's estimate, November 2, 2026, $13,858 direct, 62 lines
ChatGPT, Perplexity and paid Claude answers taken November 1, 2026; free Claude and Gemini November 2

The model, step by step

1One job, twelve runs

We used the same scope as Clamp's 5x8 bathroom: tear out and replace the 60 x 30 tub, tile the surround to the ceiling (65 sq ft), a 36 in vanity with a stone top and faucet, a toilet, an exhaust fan, two lights, five accessories, a mirror, paint, and 28 sq ft of porcelain floor. Same layout. Each run started a fresh chat with the same words and no location.

The prompt

Ask"I'm a remodeling contractor. Write me a detailed estimate for this job with line items, quantities, unit prices, labor hours and a total direct cost before my markup."

FreeFour runs, logged out

PaidThree on Sol (GPT-6.1 Sol), three on Astra (GPT-6 Astra), OpenAI's flagship

ThenOne "what did you leave out?" follow-up, and a sources-only prompt on Sol and Astra

The benchmark is Clamp's estimate for the same prompt on November 2, 2026: $13,858 direct across 62 lines, the same answer in all three runs of Clamp's free estimate generator. Permits are left out of every comparison, since a permit fee depends on the town and the prompt named none. The published 5x8 cost page shows every line.

2The totals

Every run gave a confident total. Free ChatGPT's four were 56% apart. Sol's three were 7.6% apart, and Astra's 7.1%.

Direct cost before markup and tax, vs. Clamp's $13,858
Run	Total	vs Clamp
Free 1 (GPT-6 Luna)	$9,333	−32.7%
Free 2 (GPT-6 Luna)	$11,320	−18.3%
Free 3 (GPT-6 Luna)	$13,356	−3.6%
Free 4 (GPT-6 Luna)	$14,527	+4.8%
Sol 1 (GPT-6.1 Sol)	$13,119	−5.3%
Sol 2 (GPT-6.1 Sol)	$12,193	−12.0%
Sol 3 (GPT-6.1 Sol)	$12,581	−9.2%
Astra 1 (GPT-6 Astra)	$11,909	−14.1%
Astra 2 (GPT-6 Astra)	$12,365	−10.8%
Astra 3 (GPT-6 Astra)	$12,760	−7.9%

All six paid runs added a 7% sales tax placeholder on materials and told you to replace it with your own rate; the totals here leave it out. Free run 4 printed two totals in one answer, $15,290 at the top and $14,527 in its own corrected roll-up; its lines add to $14,527.

3Free ChatGPT gets its own math wrong

Two of the four free answers have an arithmetic error a contractor would carry straight into a bid, and three guessed a city.

Errors in the free answers

Run 2Demolition subtotal $2,510; its own lines add to $2,595.50

Run 4Two totals $763 apart; a $30 a sq ft unit price on the floor tile install line on top of its labor hours

LocationRuns 2, 3 and 4 priced the job for a specific city nobody gave it, the one the browser was in

Labor rates in the free runs ran from $25 to $110 an hour, none with a source. After run 4's follow-up, free ChatGPT showed an ad for construction estimating software. Every paid line and total checked out; two Astra answers misstated a section's labor subtotal without carrying it into the total.

4Same total, different job

The paid totals land −14.1% to −5.3% under Clamp, and the money sits in different places. Every Sol and Astra run used the same three rates: $55 an hour for remodeling and tile, $85 for plumbing, $80 for electrical. At those rates it needs 114 to 128 hours. Clamp prices the same work at trade rates, 79 to 98 crew hours plus subcontracted lines, and carries a stub-out refresh, a GFCI and three devices, drywall patches and baseboard that no ChatGPT run priced.

Paid runs vs. Clamp
Item	ChatGPT	Clamp
Labor and subsChatGPT: 114 to 128 hours	$6.8k-7.7k	$8,264
Vanity packageCabinet, top, sink, faucet; ChatGPT $1,050 to $1,150	~$1,100	$2,088
TubWith setting materials	$600-710	$515
Toilet	$250-315	$350

"Stone top" read two ways: every paid run assumed a prefabricated top with the sink built in; Clamp prices a top fabricated to size. Neither is wrong. A total that lands near Clamp's but is built from different pieces is still a guess you can't check.

5The scope changes from run to run

Nobody's prompt mentioned the tub's valve or shutoffs. Each run decided for itself, so the same job came back with different work in it.

How many runs priced each item
Item	Free (4)	Paid (6)
New tub valve and trim$615 to $875 in the paid runs	0	5
Vanity shutoff valvesNo run priced the toilet's	0	5
Uncoupling membraneOthers said "underlayment"	0	1
Tile edge trimOne free run lumped it into supplies	2	6
Doorway transition	1	6

Clamp prices all five: a new valve body and trim, three shutoffs, an uncoupling membrane under the floor tile, 16 ft of edge trim and a transition. Four of the five paid runs that included the valve said how much to deduct if you keep the old one; one Astra run left it out and listed it as an add.

6Ask what it left out

After free run 4, we asked "What did you leave out of this estimate?" It found real gaps: the shower valve and trim, the toilet's shutoff valve, tile edge trim, drywall repairs and the doorway transition.

Its first total: $14,527

Flat allowance: +$1,400

Labor share of it: $500

New total: $15,927

vs Clamp: +14.9%

Gaps found, then padded

It found the gaps but fixed them with a round-number allowance instead of pricing the lines, and listed a tub drain kit as missing that its first answer already had. Use the question as a checklist, not a price.

7Ask for sources and the price disappears

The second prompt asked ChatGPT to list every scope item first, then price only what it could source. Both models wrote a better takeoff than any priced run and stopped before the prices.

Sources-only prompt
Model	Items	Priced
SolCement board and a toilet	89	2
AstraCement board and two floor tiles	102	3

Both listed the shutoffs, the flange, an uncoupling membrane, edge trim and sealant by the foot, and both made the valve a decision to settle before the walls close. Neither priced a single hour of labor. Astra ended with: "Blank prices mean unpriced, not included at no cost."

8Other assistants do the same

We ran the same prompt three times each through free Perplexity, Claude Haiku 5.5, Claude Opus 5.5 and Claude Fable 5.1, and then free Claude and free Gemini, logged out. All gave confident totals, every one under Clamp; most added a permit line, which we left out.

Direct cost before tax and permits
Assistant	vs Clamp
Perplexity, free$11,290 to $12,156; labor rates split by trade	−18.5% to −12.3%
Claude Haiku 5.5$11,933 to $13,847; kept the old valve body every time	−13.9% to −0.1%
Claude Opus 5.5$11,760 to $13,294; built a spreadsheet with formulas	−15.1% to −4.1%
Claude Fable 5.1$11,760 to $12,570; one run looked up the tub, vanity and fan	−15.1% to −9.3%
Claude, free$10,367 to $12,001; Sonnet 5.5, line lists up to 72 lines	−25.2% to −13.4%
Gemini, free$5,799 to $7,871; 3.5 Flash Lite, 8 to 13 lines, one summary that didn't add up	−58.2% to −43.2%

Each made its own call on the tub valve: one Perplexity run dropped it, all three Haiku runs swapped only the trim on the old valve body, and every Opus and Fable run replaced it. Opus caught the most scope, with shutoffs, an uncoupling membrane and a transition in every run; one Fable run added a niche nobody asked for. Eight of the nine Claude runs priced for the state the account is based in, which nobody gave them. None named a source for a labor rate; one Fable run checked current retail on the tub, vanity and fan and called the rest mid-grade averages.

9Use it for scope, not price

ChatGPT is a good scope checker: ask it for every item with quantities, then ask what it missed. The prices have to come from somewhere you can trace, your own costs or a catalog with dated rates.

Clamp's Generate from description screen with the 5x8 bathroom prompt typed in

Same prompt

The 5x8 bathroom estimate in Clamp at $13,858.38 job cost, by trade: plumbing with the stub-out update, shower valve body, trim and shutoff valves, then tile

Every line priced

The same prompt in Clamp, and the estimate it prices: $13,858.38 at job cost, every line by trade with its labor hours, including the valve, shutoffs and stub-outs the chatbots kept dropping. Illustration of the Clamp app.

Whatever writes the draft, check the scope, the quantities and the rates before markup goes on.

What it means

What to doLet ChatGPT list the scope and ask it what it missed. Don't take its prices: put every line on a rate you can trace before markup goes on.

Same job, twelve runs: $9,333 to $14,527 on free ChatGPT, $11,909 to $13,119 on paid.

Every line on a rate you can trace

Clamp turns a job description into an itemized draft at job cost, each line priced by Clamp, labor hours on every labor line, and the date its rates are from, plus what it left out.

Download on the App Store Get it on Google Play

A price you can't trace is a guess. Put the scope on your own rates.

Questions

Can ChatGPT write a construction estimate?

It writes one that looks finished, but it isn't a price you can check. In 12 runs of the same 5x8 bathroom, free ChatGPT's totals were 56% apart, and the paid models landed −14.1% to −5.3% against Clamp on labor rates they made up. Clamp prices every line from its own rates, with labor hours and the date the rates are from.

How accurate is ChatGPT at construction estimating?

On one 5x8 bathroom: free ChatGPT ran −32.7% to +4.8% against Clamp, Sol −12.0% to −5.3% and Astra −14.1% to −7.9%. The scope changed between runs, and no run named a source for a labor rate.

Is paid ChatGPT better than free for estimates?

On this job, yes: the paid runs' totals added up and stayed within a 10.2% band, −14.1% to −5.3% against Clamp. Two of four free answers had math errors, and three guessed a city. Paid didn't fix the scope drift or the untraceable rates.

Where does ChatGPT get its labor rates?

It doesn't say. Every paid run used $55, $85 and $80 an hour as "assumed" burdened costs; free runs used $25 to $110. Asked to price only from named sources, it priced no labor at all.

What's the best way to use ChatGPT for estimating?

Ask it to list every scope item with quantities before any price, then ask what it left out. Price the list from your own costs or a catalog with dated rates.

Sources & provenance

12 ChatGPT runs of one 5x8 bathroom estimate Clamp Research
Clamp's estimate for the 5x8 bathroom (bath-5x8/v5, 2026-11-02) Clamp Research
OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less TechCrunch (research)
USG Durock EdgeGuard 1/2 in x 3 ft x 5 ft cement board (172954) Home Depot (research)

Changes: Nov 1, 2026: Published: 12 ChatGPT runs of one 5x8 bathroom, plus a follow-up, checked against Clamp's price.; Nov 2, 2026: Rebased against bath-5x8/v5, the 5x8 bathroom as Clamp's free estimate generator prices it ($13,858.38, 62 lines); added free Claude and free Gemini rows.; Nov 2, 2026: Run labels in the totals table name the model.; Nov 2, 2026: Removed the city, county and state names the chatbots guessed; the guides say only that the guess matched where the account is based.