Asking minified JSON can reduce output token costs by 30% (getsnapkin.app)

Recent findings reveal that instructing AI models to generate minified JSON outputs can reduce token costs by up to 30% while also speeding up response times. In a study involving ten AI models, seven exhibited significant savings by transitioning from pretty-printed JSON, which includes line breaks and indentation, to a compact single-line format with just one additional instruction in the prompt. This change primarily impacts the output phase, leading to faster generation times, particularly for models that begin writing immediately upon processing input. The significance of this finding is twofold for the AI/ML community. Firstly, it presents a straightforward method for optimizing the efficiency and cost-effectiveness of model responses, especially in scenarios where output token consumption is high. Secondly, understanding that certain AI models already output minified JSON can avoid unnecessary prompts, which can hinder performance. The research encourages users to evaluate their models for output format and consider implementing minified JSON instructions to maximize both speed and cost savings in AI applications.