tokensitter

What each lever saves.

Lab results: fixed tasks and measurements, run with a lever off and on or at two levels, costs at Anthropic API list prices. They show what a lever can do on those tasks, not what it will do on your bill.

Before you quote a number.

  • These are fixed tasks run in a test harness, not your sessions. Your saving depends on your prompts, your files and how long your sessions run.
  • Short replies and lean code ran as cold sessions with a small synthetic cached prefix. Real Claude Code prefixes are larger; with a bigger prefix the percentage is smaller, with a warm cache it is larger.
  • The coding tasks are small: every level passed every task, so a quality cost would not show at this size.
  • Ultra's weak spot is a new question late in a conversation, where a terse reply can drop a detail.
  • Read less was tested on synthetic files. Its hooks fired in 10 of 54 sessions with it on, so the total mixes those with sessions where on and off behave alike.
  • Costs are at Anthropic API list prices. On a Pro or Max plan they show how fast you use your limits, not your invoice.
  • Cache: the one-hour expiry was not waited out; the cold resume is a never-seen prefix, which bills the same full cache write.
  • Cache: how often people come back after an hour or more was not measured; the break-even shares say how often they must.
  • Cheaper helpers: the tasks are easy read-only lookups, so this shows Haiku is enough there, not on hard reasoning or edits.
  • Cheaper helpers: the answer key was widened after the 3 misses, so the as-run score is 77/80 and the audited one 80/80.
  • Cheaper helpers: Haiku was forced on every task; how often the product's classifier picks right was not measured.
  • MCPs & Skills: the plugins are the public official set and the three setups are illustrative bundles, not a survey of real installs.
  • MCPs & Skills: this is the cost of carrying the context, not the benefit of a skill or MCP you use.