What a design costs in tokens
15 Nov 2026. Behind the card on the home page.
Three starter-kit screens that come with the app (Home, Settings, Sign in), each written once as AgentDesign calls and once as a single HTML file. Every figure is a count of tokens with tiktoken 0.14.0 and its o200k_base encoding, OpenAI's public tokenizer: none is a Claude count. A picture counts as ⌈width / 28⌉ × ⌈height / 28⌉ tokens (465 for these screens). The build is a release build from 15 November 2026, run on a Mac with its fonts, twice from nothing, with identical counts.
Three cases follow the client your agent uses: it lists tool names only, it loads the tools it uses, or it loads all 80 tool definitions at once. Where a figure has three values, they are in that order.
Per call
- Writing a screen
- 1,987 tokens in text for the three kit screens, against 4,698 as HTML: 42% overall, 40 to 47% a screen. With the picture each write returns, 3,382 tokens: 72%.
- Reading one back
- 45 to 53% of the HTML.
- Changing one text
- 164 to 173 tokens in text, 239 to 248 with its picture, against 50 to 54 for a diff edit of HTML.
Once a session
- Instructions
- 455 tokens.
- Schema call and reply
- 2,464 tokens.
- Tool definitions
- 262 / 2,164 / 20,211 tokens: names only, the 7 a first session uses, all 80.
- All of it
- 3,181 tokens when the client lists names only, up to 23,130 when it loads all 80 definitions.
Whole jobs
- First session, three kit screens
- 6,563 / 8,465 / 26,512 tokens with the pictures (the first is 5,168 in text only), against 4,698 for HTML (6,093 if the HTML agent also looks at each screen). HTML is lower.
- Later session, one text changed on each of three screens
- 5,834 / 6,358 / 25,783 tokens with the pictures, against 4,210 for HTML (5,605 if it looks). HTML is lower.
- Pays back, text only
- 3.52 / 5.62 / 25.60 new screens in one session: the fixed cost (3,181 / 5,083 / 23,130) divided by the 903.67 tokens a screen saves (2,711 over three screens).
- Pays back, with each write's picture
- 7.25 / 11.59 / 52.73 new screens in one session, against an HTML agent that does not look: the same fixed cost divided by the 438.67 tokens a screen saves (1,316 over three screens).
What this does not show
- One agent wrote the HTML in one pass; another would write more or less.
- The replies to the HTML agent's calls are not counted, which favors HTML.
- Thinking, retries and prompt caching are not counted on either side.
- Three phone-sized screens is a small sample, and these are OpenAI token counts, not Claude's.
- The same build run on a machine without the Mac's system fonts gets about 20 font warnings in every write reply. Taken as printed, it breaks even after 11.8 / 18.8 / 85.5 screens in text and never with the picture. That is a property of the machine, so it is not the measurement; it does mean an agent on such a machine pays those warnings on every write.