Preterview cuts Claude token costs 58% after fixing prompt caching bugs
Tactic · Dev.to · stat: 94% hits Preterview founder Ji AI reported a 0% cache hit rate across 1,412 consecutive Claude API calls due to three subtle implementation bugs, including a dynamic system…
Tactic · Dev.to · stat: 94% hits
Preterview founder Ji AI reported a 0% cache hit rate across 1,412 consecutive Claude API calls due to three subtle implementation bugs, including a dynamic system prompt timestamp. Sorting tool definitions, removing timestamps, and enforcing minimum prompt lengths boosted cache hits to 94% and lowered input token costs by 58%.
Unmonitored LLM prompt caches routinely leak margin without returning errors Developers using Anthropic prompt caching must explicitly log token usage fields to ensure dynamic prefixes are not triggering silent full-price cache misses.
Every claim ties to a primary source. See our methodology.