Check our savings on your own bill

Don't take our counters for it. Run the same tasks twice, once with Tokensitter off and once on, and let Anthropic's bill decide. Setting up takes about 10 minutes; the runs take as long as your tasks.

Why the bill, not reply length

Counting how much shorter the replies got only shows part of the cost. Your bill also has every message's input, cache writes and cache reads, and the work of helper models. Shorter replies add a small instruction to every message, and cheaper helpers move work to a cheaper model: only the whole bill shows both.

The test

Two workspaces. In the Claude Console, create two workspaces, for example ts-off and ts-on, and one API key in each. Each run then has its own line on the Cost page.
Pick your tasks. Choose 5 to 10 tasks you really do. Write down the first message of each, word for word, so both runs start the same. Use the same model and effort for both runs.
Run A, Tokensitter off. In Claude Code, open /plugin and disable runway@tokensitter. Start Claude Code with the ts-off key, run each task in a fresh session, and stop when it is done.
Run B, Tokensitter on. Enable runway@tokensitter again, start Claude Code with the ts-on key, and run the same tasks the same way.
Compare. On the Console's Cost page, compare the two workspaces. The Usage page shows the same split by API key and model, in tokens.

Read it fairly

Models never answer the same way twice, so two runs differ even with nothing changed. More tasks give a steadier answer; if the gap is small, run both again.
Cheaper helpers only act on helper tasks that only read. If your tasks never hand work to a helper, that lever has nothing to save here.
Keeping the cache warm is off by default and only acts after an hour away, so a short test will not show it.
The test needs API billing, and it costs what your tasks cost, twice. On a Claude plan (Pro or Max) there is no bill per workspace.

What our lab measured

For comparison, on fixed tasks: shorter replies cut the total cost of a Sonnet chat conversation by 25% and of an Opus coding task by 4%; cheaper helpers cut the total cost of a read-only lookup task by 70.9%. How each was measured is on Benchmarks. If your bill disagrees with us, trust your bill, and tell us: we would like to know.