How Snapkin is built

The ideas behind a simple-looking app, each shown in a figure. The experiments that did not work belong here too.

AI speed and cost · Latest

Your LLM is billing you for whitespace

On a real production request, seven of ten AI models pretty-printed their JSON. Asking for minified JSON cuts latency and cost on those.

Mattias Geniar 15 Oct 2026 4 min read

Indents its JSON 7 Already one line 3

Gemini 4 Argon

Normal

30.5 s

Minified

−9.0 s

21.4 s

−30%

Mistral Large 4thinking switched off

Normal

25.6 s

Minified

−6.4 s

19.3 s

−25%

Gemini 3.8 Flash

Normal

11.4 s

Minified

−2.5 s

8.8 s

−22%

GPT-6.1

Normal

24.4 s

Minified

−5.2 s

19.2 s

−21%

GLM 5.3 Flashindents sometimes

Normal

12.5 s

Minified

11.3 s

−9%

Qwen 3.8 Maxthinks first, indents sometimes

Normal

86.9 s

Minified

80.5 s

−7%

DeepSeek V4.1 Flashthinks first

Normal

27.9 s

Minified

26.0 s

−7%

Claude Opus 5.5

Normal

25.7 s

Minified

25.7 s

Kimi K3

Normal

9.5 s

Minified

9.5 s

Grok 4.7thinks first

Normal

48.0 s

Minified

48.0 s

Reading the inputThinkingWriting the answer The difference

Seconds until the whole answer is in, with and without minified JSON. The dashed box is the time saved. Each model has its own scale.