Fig 1cachellmHover to weigh · 1 2 3 set the modelocked 1-4 · 11,800 tok · 4 of 4 set
The segment being weighed, then the locked prefix that liftsNine segments of an example prompt: length is estimated tokens, reach across the galley is the stability scoreTwo segments under 1,024 estimated tokens, which take no breakpointcachellm wraps an Anthropic, OpenAI or Gemini client you already have. For an Anthropic request it splits the prompt into system, tools and message segments, scores each one for how likely it is to repeat, passes over anything under 1,024 estimated tokens, and places up to four cache breakpoints by stability times size, so everything before the last one is reused on the next call instead of being paid for again.