Tag / prompt caching

Anthropic's input_tokens excludes cached tokens, OpenAI's includes them. Port a cost function between the two and a $0.082 request bills $0.378. Fields, traps, code.

The model 5.7% cheaper on the price tag billed an agent session 4.4x higher. The cached-read rate decides an agent bill, and it runs from 2% to 25% of input.

Every extra agent pays for its own cold prefix. But re-reading history grows with the square of turns, so splitting one job across three agents ran 17% cheaper.

A cache miss costs exactly 20x the warm read of the same history. What that does to your plan limits, which settings change it, and the number nobody can compute.

Batch halves token rates at OpenAI, Anthropic and Google -- but takes 20% at SpaceXAI and nothing on its flagship. When batch beats a warm cache, and when it costs more.

Prompt caching fails silently -- no error, no warning, just a full-price bill. The floors, the breakpoint rules, and everything that quietly invalidates a cache.

Ten routers launched on HN in five weeks. The peer-reviewed 2026 benchmarks say they barely beat a baseline. The arithmetic that decides it, on verified rates.

Five levers, one deployment order: which context feature to reach for by symptom, what each does to your cache and your bill, and one config that runs them all.

Cutting 63% of an agent's tokens moved its bill 9%. Cutting 1.2% moved it 15%. The arithmetic that converts token reductions into dollars, on verified 2026 rates.

Five MCP servers burn ~55k tokens before you ask anything. Tool search and programmatic tool calling (both now GA) cut that 85%+ — with one caveat that bites.

Anthropic's compaction API summarizes an agent's history when it hits a token threshold. How it works, the billing pass you don't see, and when it backfires.