What does a design actually cost?

How much does it cost to get a design out of Claude? Not per seat or per month, but per thing: a new page, a change to one, a restyle. That’s the number a designer can put next to the hours the job would have taken, and I’ve hardly ever seen anyone publish it.

So I measured it.

How I measured it

Claude Design runs on a subscription, and it doesn’t show you tokens. So I ran the same kinds of job through Claude Code on Opus 5.5, where every call logs exactly what went in and out, and priced them at Anthropic’s list API rates for 9 November 2026 (in US dollars). On a subscription you pay a flat fee and these jobs come out of your usage limits instead. The shape of the cost is the same either way.

The site is Gritline again, the made-up climbing club from my last post. Three new jobs started from the pages built for that post, and the new-page numbers come from the three builds in it. Each job is one run, in a fresh agent with no skills loaded except where I say so, so this is a demonstration rather than a benchmark. Run it again and the numbers will move a bit.

Six runs, priced

Outcome	The brief	Time	Cost
Change the copy	Rewrite the hero to sell a winter beginners’ course	36 seconds	$0.44
Modify the site	Add a “first day on the grit” section and a beginner filter on the guidebook	1 minute	$0.62
Restyle the design	A new identity: palette, fonts, a light theme and contrast checks	3½ minutes	$0.98
New page, no skill	A whole site from a written brief	3½ minutes	$0.95
New page, stock skill	The same brief with Anthropic’s frontend-design skill	5½ minutes	$1.47
New page, my skill	The same brief with my modified skill	6½ minutes	$1.37

Every one of them came in under $1.50. Honestly, I expected the spread to be wider. A full restyle of a page costs about the same as building one, and changing a headline still costs more than 40 cents.

Every job pays an entry fee

The copy change explains why. Before an agent reads anything of yours, its first call loads the harness: Claude Code’s own instructions and the definitions of every tool it can use. That came to about 57,000 tokens in every run, and writing it to the cache costs 29 cents at Opus 5.5’s rate of $5 per million tokens. The copy change only needed to read the hero, and the new words it wrote cost 5 cents, so most of its 44 cents went on turning up.

So a one-line change costs nearly as much as a new section, because both pay the same fee. If you want small changes to be cheap, batch them: one agent with ten tweaks to make pays that fee once, and ten agents with one tweak each pay it ten times.

The skills show the other side of it. The fee is paid once, but every call after it rereads the conversation so far. Without a skill, the new page took 5 calls to the model. With either skill it took 16 to 18, because the skill makes the agent plan, check the plan against the defaults and then build. That’s why the skill runs read about 1.5 million cached tokens against 310,000 for the run without one. Cached tokens are cheap at 20 cents per million, but the planning also meant more output, and together that added about 50 cents a page.

Was it worth 50 cents? Look at the side-by-side screenshots and decide for yourself. I’d pay it every time.

A build isn’t an outcome

Every one of those runs handed back a page that looked finished, and every one needed checking before anyone could call it done.

• The copy change rewired the hero buttons. Nobody asked it to, and its reason was good (the old buttons pointed at sections that had nothing to do with the course), but it was still a decision someone had to notice.
• The new section told newcomers which bus to catch to Stanage, and the agent said plainly that the route, journey times and last bus were its own assumptions. A beginner would trust that page.
• The restyle passed its contrast checks in both themes. It also quietly changed what the colours meant: lichen green used to mark damp rock, and now heather, chalk and grey do that job. That’s a design change, not a restyle.
• None of the three jobs opened its page in a browser, so the guidebook toggle and the phone layout are untested.

Checking that is what the AI-DLC (the AI-driven development lifecycle) is for. In the outcome post I argued that work isn’t done until each step has handed back evidence at its gate. A page that looks right is the build step reporting in. “The bus times are real, the colours still mean what they meant, it works on a phone” is the gate, and right now the gate is a designer’s time.

It’s also the same argument as four metrics, then outcomes. Tokens are an activity metric. Cost per outcome is the first number that attaches to something the business actually asked for, but only if you count what it took to close it.

The whole receipt

So what did a whole outcome cost? My last post is the nearest thing I have to a complete one: three builds compared, screenshotted, written up, reviewed and published.

Pick a skill, make it yours

8 Nov 2026 · 21:17 to 22:53 · Opus 5.5

Build

Run A, no skill

$0.95

Run B, stock skill

$1.47

Run C, my skill

$1.37

Gates

Voice reviews and preflight checks (17 agents)

$1.16

Research on the skills shelf

$0.02

Direction

Briefs, screenshots, writing and fixes

$15.00

Total

$19.97

Builds were 19% of the bill.

The builds were less than a fifth of it. The automated gates were cheap too: each voice review on Sonnet 5.5 cost between 6 and 12 cents. Three-quarters of the money went on the long session where I directed the work, took the screenshots, wrote the post and fixed what the reviewers found.

Most of that was rereading. That stretch of the session ran for over an hour and a half, and every turn read the whole conversation back in, which came to 54 million cached tokens and $10.78 of the $15. So if you’re budgeting for this, hand the builds to short, fresh agents, and keep the long session for decisions.

And the $20 doesn’t include me. I was there for the whole hour and a half, and that’s the number to put against it.

Final thought

The tokens for a design job cost about the same as a coffee, and that’s true whether it’s a headline or a whole site. What makes a design expensive is everything around the build: the fee for starting every job, the turns spent deciding, and the person who checks the bus times. So next time you ask what AI costs your design team, pick one finished outcome and add up the build, the gates and the hours of the person who closed it.