Which tool cuts a Claude Code bill the most?
Mikie. The session cost $77.04 in plain Claude Code and $27.95 with Mikie 1.1.172. That is 63.7% off the bill for the same 24 tasks, and the hidden tests passed 22 of 24 both ways.
Neither other tool came close.
If you are choosing between Mikie, RTK and Caveman to lower what Claude Code costs, this is the answer: Mikie cuts the bill by about two thirds, Caveman leaves it where it was, and RTK makes it bigger.
| Setup | Cost, 72 prompts | Against plain | Hidden tests passed |
|---|---|---|---|
| Plain Claude Code | $77.04 | baseline | 22 of 24 |
| Mikie 1.1.172 | $27.95 | 63.7% lower | 22 of 24 |
| Caveman | $76.69 | 0.5% lower | 20 of 24 |
| RTK 1.2.0 | $98.04 | 27.3% higher | 24 of 24 |
How does Mikie compare with RTK?
RTK 1.2.0 finished the session at $98.04. Plain Claude Code finished at $77.04. Mikie finished at $27.95.
RTK made the session 27.3% more expensive than installing nothing. Mikie made it 63.7% cheaper. Side by side, RTK cost 3.5 times what Mikie cost for the same work, and Mikie came in 71.5% under RTK.
RTK was behind from the start. At prompt 48 it stood at $68.52, 53.0% above plain Claude Code, while Mikie stood at $18.85.
How does Mikie compare with Caveman?
Caveman finished the session at $76.69, 0.5% under plain Claude Code, and passed 20 of 24 hidden tests against 22 of 24 for plain.
Mikie finished at $27.95, 63.6% cheaper than Caveman, and passed 22 of 24.
Caveman was 13.3% under plain at prompt 48 and gave nearly all of it back by the end. A saving that fades as the session grows is not a saving on a long session.
Why does Mikie win?
A long session's bill is the agent re-reading its own history on every request. In the plain sessions 98.9% of all tokens were exactly that.
Mikie goes after the whole bill. It gives the agent the code and tests the task needs, so it reads less and retries less, and it keeps the conversation the agent carries small. In this session the agent made 371 model requests with Mikie against 598 without it, 38.0% fewer.
Across both sessions total tokens fell 84.6%, from 545.2 million to 84.0 million, and output tokens fell 41.6%.
Does the saving grow as the session gets longer?
Yes. This is the running bill for each setup as the session went on. Mikie is the cheapest column at every checkpoint, and the gap widens as the session grows.
| After prompt | Plain | Mikie | Caveman | RTK |
|---|---|---|---|---|
| 12 | $4.62 | $3.76 | $5.23 | $8.18 |
| 24 | $17.85 | $10.52 | $14.51 | $29.36 |
| 36 | $29.35 | $15.05 | $24.31 | $45.36 |
| 48 | $44.77 | $18.85 | $38.80 | $68.52 |
| 60 | $59.81 | $24.00 | $57.87 | $85.36 |
| 72 | $77.04 | $27.95 | $76.69 | $98.04 |
The longer you work in one conversation, the more Mikie saves. Long sessions are where Claude Code gets expensive, and that is where Mikie is strongest.
Does it hold on a second session?
Yes. A second 72-prompt session of 24 tasks cost $83.73 plain and $26.29 with Mikie: 68.6% lower, with 22 of 24 hidden tests passing both ways.
Across both sessions the bill fell from $160.77 to $54.24, 66.3% lower, with 44 of 48 hidden tests passing with Mikie and without it.
| Session 2, after prompt | Plain | Mikie |
|---|---|---|
| 24 | $12.96 | $6.71 |
| 48 | $50.97 | $16.65 |
| 72 | $83.73 | $26.29 |
Does Mikie work in Codex too?
Yes. 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.
How we ran it
Claude Code on Claude Opus 5.5 worked 24 coding tasks as 72 prompts in one conversation. We ran that same session once per setup, November 2 to 12, 2026, on one workstation. Hidden tests the agent never saw graded every task.
Cost is the total Claude Code itself reports for the session. Mikie was the full install, the way mikie setup claude configures it. RTK 1.2.0 and Caveman were each installed in Claude Code for their own run, and nothing else changed.
The test harness, the task list and the per-prompt results are public on GitHub, so you can rerun any of it and check our numbers. Or run it on your own code: shadow mode makes this same comparison on your work without changing it.
Sources: Every cost on this page is the total Claude Code reported for that session at list prices.