Building Snapkin means running a lot of experiments. Which AI model reads a plate best, what a prompt change does to the accuracy, how long you wait for an answer, what each call costs, …
Those results get a home of their own now: Snapkin Engineering .
Every post is one idea, built around a figure. The experiments that didn’t work go on there too, because those are often the more useful ones to read. (There’s also a research section that covers the nutrition science behind the app. The engineering blog is the other half: how it’s built.)
Your LLM is billing you for whitespace# #
The first post is about JSON. Snapkin asks an AI model to describe a meal as structured output: the dish, the items on the plate, the calories, and the follow-up questions it wants to ask you.
Turns out most models pretty-print that JSON. Every field gets its own line, every line gets indented. You pay for each of those spaces and newlines as output tokens, and you wait while the model writes them, one token at a time.
I sent the same meal photo, prompt and JSON schema to ten models. On seven of them, a big chunk of the answer was whitespace:
Five models indent every answer, and on those, 27 to 34% of the output tokens are spaces and line breaks. Claude Opus 5.5, Kimi K3 and Grok 4.7 already write single-line JSON.
The fix is one sentence, appended to the system prompt:
Write the JSON minified: one line, no spaces or line breaks outside string values.
On the models that indent, the whole call got up to 30% faster (Gemini 4 Argon went from 30.5 to 21.4 seconds) and up to 28% cheaper:
The gain shrinks on reasoning models. Thinking is billed as output too and minifying doesn’t touch it, so DeepSeek and Qwen only save 5 to 9%. On the three models that already write one line, there’s nothing to gain at all. Grok even thought longer with the instruction, so leave it out there.
Whether a model indents also depends on the request, not just the model. On a tiny test prompt, Gemini 3.8 Flash, GPT-6.1 and DeepSeek wrote single-line JSON without being asked. On the real production prompt, with its long system prompt and large schema, they indented. A quick test in a playground tells you very little: look at the raw output of your actual request.
The full post has the per-model numbers, the latency breakdown and how it was measured. Snapkin now asks for minified JSON on every meal analysis, and on every correction you make to one.
Following along# #
There’s an Atom feed if you want new posts in your feed reader. No newsletter, no tracking.
If you run structured output in production and you’ve measured something similar (or the opposite), let me know. I’d love to compare notes.