When Tokensitter won't save you money

Every lever has a cost or a limit. These are the cases we know of, from our own lab tests. Where we have not measured something, we say so.

Very short questions

Shorter replies adds one short instruction to every message you send. If your answers are already short, there is little left to cut, and you still send the instruction each time. It pays where replies run long: in our lab it cut the total cost of a Sonnet chat conversation by 25%. On Opus coding tasks the cut was 4%, because most of that output is code, which we never shorten.

A flat plan you never max out

On Pro or Max, if you never hit the limit and never buy extra usage, fewer tokens give you headroom, not cash. Your bill to Anthropic stays the same, and so that month costs $0 with us. See how pricing works.

Keeping the cache warm for sessions nobody resumes

This lever is off by default. When on, it sends a one-line ping to keep a long session's cache alive while you are away. On a session of about 100,000 tokens each ping costs $0.0204. If you never come back, that money buys nothing: $0.0204, $0.0408 or $0.0613 for 1, 2 or 3 pings. When you do come back after an hour, it saves $0.3597 per resume. It pays off only if at least 5% of idle sessions end in a resume after an hour (11% after 2 hours, 16% after 3). Sessions on the 5-minute cache are never pinged.

Hard work on cheaper helpers

We tested cheaper helpers only on easy, read-only lookups in a small codebase. There, a Haiku helper cut the total cost per task by 70.9%, with the same correct answers. We have not tested hard reasoning, edits or unclear tasks. Tokensitter only moves helper tasks it judges clearly read-only, and in a long session the main model does more of the work, so the share saved is smaller than in the lab.

MCP servers and skills you use

We measure what always-loaded tools cost in every message, not what they are worth. A server or skill you use can save far more than it costs. Switch off only what you have not used.

Plan usage is not cash

Our lab figures are at Anthropic's API list prices. On a subscription plan the same savings are plan usage, not money, until they keep you from buying extra usage or let you move to a cheaper plan.