Your LLM is billing you for whitespace (getsnapkin.app)

A recent analysis highlights a cost-saving opportunity for users of large language models (LLMs) when generating structured outputs, particularly in JSON format. By instructing models to deliver minified JSON—condensed into a single line without unnecessary whitespace—users can achieve significant savings in both time and expenditure. The study examined ten models and found that for seven of them, including Gemini 4 Argon and Mistral Large, switching to minified output can reduce response latency by up to 30% and overall costs by 25-28%. This finding is crucial for developers and businesses relying on LLMs, as it emphasizes the importance of tailored output formats to optimize performance and costs. Notably, while some models already generate a minified version without prompting, others may still default to pretty-printed output, leading to increased token usage and billing. The implications are clear: users should assess their model's output style and consider implementing prompts for minified responses to maximize efficiency and minimize operational costs.