Glossary · term144 of 202
prompt caching
A billing mechanism where a stable, byte-identical prompt prefix gets reused across turns at a fraction of normal input cost, while any edit to that prefix forces a full-price rewrite.
also written as cached prefix
Related terms
cache hit ratioThe share of total tokens in a session that came from cache reads rather than fresh input, which matters because cached tokens bill far cheaper than uncached ones.view termcache missA request whose prompt prefix no longer matches what was cached, forcing the full prefix to be billed again at the cache-write rate instead of the cheaper cache-read rate.view termcache readA token served from a prior prompt's cached context instead of being processed fresh, billed at a fraction of the base input price. Most tokens in a long agent session are cache reads, not new context.view termcache writeThe cost of storing a prompt's context so later requests can read it from cache instead of reprocessing it. Billed above base input price, with the rate depending on how long the write stays valid.view termcompactionA billed request that reads an entire conversation history and replaces it with a shorter summary. The read itself counts as full input tokens, so compacting often costs more upfront than it saves unless the summary gets reused many times.view termeffortA per-subagent setting that controls how much a model deliberates before acting, trading thoroughness for speed, fewer tool calls, and lower spend. Left unset, it silently inherits the session's higher default.view term
At a glance
- cited by
- 5 posts
- categories
- 3
- first used
- apr 2026
Why expensive models make terrible execution subagentsSubagent model selection has one costly default. Reasoning models obey their own traces 15.6% of the time, so here is where to pin model and effort.
Token-Compression Skills Do Not Survive MeasurementToken compression skills report savings the invoice never shows. Cache reads carry 87 percent of an agent bill, and cutting tokens can push the cost up.
Ways to Make the Most Out of Claude Fable 5 (2026)Get more from Claude Fable 5: pick the right effort level, plan with Fable while cheaper models implement, and claw back costs with caching and batching.
AI Agents in Production: What Actually Works After 6 MonthsMost AI agent projects die between demo and production. After 6 months running agents at scale, these are the patterns that survived: model routing, trace-based debugging, loop detection, and hard cost ceilings.
Claude Usage Limits Are a Bigger Problem Than Anthropic AdmitsClaude usage limits burned $200 plans in 19 minutes. Zero published quotas, peak-hour rationing, and opacity are driving developers away from Anthropic.