THE HUNGRY, HUNGRY AI MODEL

Summary from the TLDR Founders newsletter

AI models consume 300x more input tokens than they output - sometimes up to 4,000x. This founder tested it expecting 20x, but discovered the shocking reality. Since input tokens cost 98% of your API bill despite being 4x cheaper per token, the engineering challenge isn't prompting anymore - it's context engineering. You need efficient data retrieval that finds the best information in the smallest token footprint. Caching moves from nice-to-have to mission-critical when 99% of your tokens are input.