Surprisingly good design judgement

Fast & cost effective for generating visuals from short prompts.

We used Sonnet as part of a two-pass LLM pipeline for taking a short user-provided prompt describing a graphic (e.g. "lifecycle of a cup of coffee" or "project timeline with three lanes over two years, segmented by quarter") and producing an SVG depicting the result. The first phase is taking the prompt and describing the desired SVG including typography choices, colors, and text content; the second phase is taking that description and producing the SVG.

For cost effectiveness we started with open-weight models for both phases, but many of those gave poor quality visual results and invalid SVG. Sonnet, while more expensive, was the sweet spot in terms of coming up with good design ideas and expressing them in a way that a cheaper model could be used to generate the actual SVG (the second phase). So, we ended up shipping Sonnet as the model for the first phase. Anecdotally, pricing ended up around ~$0.01 or slightly less per request.

Sonnet does produce SVG with invalid syntax maybe 20% of the time, but we addressed that by invoking an optional third phase of syntax fix-up using a super cheap model.