Is Mikie a token optimizer?
Mikie reduces tokens. So does every tool in this corner of the market, and that is the problem with the label: it names one input and stops there.
Mikie is cost reduction for coding agents. The measure is the bill for the same work, with the same tests passing. Tokens are reported beside it, because they explain part of the bill. They are not the bill.
Why is a token cut different from a cost cut?
Tokens are priced by kind. Output costs the most. Fresh input costs less. Input the provider has already cached costs a small fraction of that.
We ran 2 Claude Code sessions of 24 tasks each on Claude Opus 5.5, once plain and once with Mikie 1.1.172. In the plain sessions, 98.9% of all tokens were cached reads: the agent re-reading its own history on each request.
Mikie cut total tokens 84.6%, from 545.2 million to 84.0 million. Cost fell 66.3%, from $160.77 to $54.24. The token number is bigger because most of the tokens removed were the lowest-priced kind.
A tool that reports only the token number is reporting the flattering one.
Can a token-saving tool raise your bill?
Yes. A model that is missing what it needs does not stop. It reads again, searches again, and retries. Every extra request re-sends the conversation so far, and every retry is paid for.
So a tool can shorten each prompt, show fewer tokens per request, and still leave you with more requests and a larger total. The saving per request is real. The bill went the other way.
It happened to us. In one short session Mikie cost $1.77 against $1.06 without it, and both fixes passed. We publish that session with the rest.
What did Mikie do to the bill?
Across the two long sessions the bill fell 66.3%. Hidden tests passed 44 of 48 without Mikie and 44 of 48 with it. Model requests fell 40.9%, from 1,248 to 738, and output tokens fell 41.6%. Fewer requests is the opposite of a retry loop.
On 36 held-out Codex tasks, cost per run fell from $8.50 to $3.36, 60.4% lower, with every counted fix accepted both ways.
Where does Mikie save less?
Short sessions. 8 three-prompt Claude Code sessions cost $11.71 without Mikie and $8.28 with it, 29.2% lower. A short session has little history to re-read, so there is less to remove.
The long-session result is two sessions on one model and one workstation, and each condition ran once. The plain sessions and the Mikie sessions ran a week apart. The method page lists every limit.
How do you check a cost claim?
Ask any tool, including this one, for three things together:
- The cost with the tool and without it, in money.
- On the same tasks, with the same model and settings.
- With the tests passing in both.
A token count answers none of the three. A receipt for the same work answers all of them. Shadow mode runs that comparison on your own code without changing it.
Sources: Long-session figures: 2 Claude Code sessions of 24 tasks each on Claude Opus 5.5, plain against Mikie 1.1.172, measured November 10, 2026. Tokens are provider-reported; cost is the total Claude Code reports for each session at list prices. Held-out figures: a held-out benchmark of 36 tasks in Python, TypeScript and Rust (2026-10-07). Method and limits are on the evidence policy page.