I Cut Vercel's JSON-Render LLM Costs by 89% Using Toon (mateolafalce.github.io)

In a recent blog post, Mateo Lafalce revealed how switching from JSONL to the TOON output format significantly reduced the costs of Vercel's dynamic UI generation tool, json-render, by 89%. The original implementation, which utilized the Claude Opus 4.5 model, was costly due to JSONL's verbosity—output tokens were three times more expensive than input tokens. Lafalce's hypothesis hinged on the idea that a more compact format like TOON could lower the output token count enough to offset any additional input token requirements. The benchmarking results not only confirmed substantial improvements in token efficiency, cost savings, and performance but also highlighted a crucial limitation: TOON does not support streaming responses as JSONL does, which means the full output must be generated before use. This finding emphasizes a vital lesson for the AI/ML community: optimizing output formats is essential, especially when output tokens are pricier. Lafalce’s approach serves as a blueprint for developers aiming to build cost-effective AI applications, underscoring the importance of rethinking output strategies in LLM usage.