Mouse-1 makes the same agent finish 4.7× faster and about 7× cheaper than an Opus 5.5 judge: 18.1 s and $0.025 per task against 84.2 s and $0.1737, median of five runs. Every run in both arms passed the hidden tests.
Some harnesses route their judgment calls (is this safe, is this done, which tools) to the strongest model available, on the theory that a smarter judge makes fewer mistakes. Claude Opus 5.5 is that model for many teams, at $4 per 1M input tokens and $20 per 1M output. We raced it against Mouse-1 on the same task.
The race
The same Claude Sonnet 5.5 agent fixed the same planted bug (expired API keys still authenticate) twice per pair, five pairs, interleaved. Code-writing turns were identical. Only the harness's decisions changed hands: a gate on every bash call, a completion check when the agent stopped, and, in the Mouse arm, choosing which of 50 connected tools to load.
| Median of 5 runs | Opus 5.5 judge | Mouse-1 |
|---|---|---|
| Wall time per task | 84.2 s | 18.1 s |
| Time waiting on decisions | 43.6 s | 4.4 s |
| Total cost per task | $0.1737 | $0.025 |
| Cost of the decisions alone | $0.08231 | $0.00045 |
| Mean cost per successful task | $0.1446 | $0.0271 |
| Pass rate | 100% | 100% |
Median of 5 runs, seconds · Lower is better
| Mouse-1 | 18.1 s |
|---|---|
| Opus 5.5 judge | 84.2 s |
Where the time and money go
Each verdict is slow. Opus reasons before it answers. On a single yes/no gate sent 20 times to each from the same machine, the median round trip was 344 ms for Mouse-1 and 3,112 ms for Opus, even at low effort. The race ran Opus at medium effort.
A strict judge makes the agent keep working. In 4 of 5 runs, the agent reported a working fix and Opus answered "not done" at the first completion check. The agent then added regression tests and asked again, sometimes twice more. That doubled the agent's turns (a median of 11 frontier calls against 5). Of the cost gap, a little over half is the Opus decision calls themselves and the rest is those extra agent turns. Every Mouse-1 run accepted the fix at its first check, and every run in both arms passed the hidden grader.
Whether that strictness is worth paying for depends on your bar. If "done" should mean "done with regression tests", tell the completion check so, in its instructions, and Mouse-1 will judge against that bar in a few hundred milliseconds.
How we measured
- Traces for every run are in
bench/race/runs/2026-10-20-opus-judge-n5(bench/race/race.py). The agent was Sonnet 5.5 at medium effort; the Opus judge ran at medium effort through OpenRouter with a strict JSON schema. - Single-gate latency:
bench/race/latency.py, 20 interleaved calls per model from one laptop, all LLM judges at low effort. - Five pairs on one task is a small sample. Read the medians as direction, not precision.
Sources
Checked October 20, 2026.
See also: Mouse-1 vs Claude Haiku as a judge · Mouse-1 vs GPT-6 Sol · All comparisons
FAQ
Is Mouse-1 faster than Claude Opus 5.5 as an agent judge?
Yes. With the same Sonnet 5.5 agent fixing the same bug, the median task took 18.1 seconds with Mouse-1 answering its decisions against 84.2 seconds with an Opus 5.5 judge, and spent 4.4 seconds waiting on decisions against 43.6. On a single yes/no gate, Mouse-1's median round trip was 344 ms against 3,112 ms for Opus.
How much cheaper is Mouse-1 than an Opus 5.5 judge?
In our runs the median task cost $0.025 with Mouse-1 and $0.174 with an Opus 5.5 judge, about 7 times less. The decisions alone cost $0.00045 against $0.082. Opus 5.5 is priced at $4 per 1M input tokens and $20 per 1M output; Mouse-1 at $0.10 per 1M input with free output.
Is an Opus judge more careful than Mouse-1?
It is stricter. In four of five runs Opus answered 'not done' at the first completion check after the agent reported a working fix, and the agent kept going. All ten runs passed the hidden grader, including every Mouse-1 run, which accepted at its first check, so the extra turns did not change the outcome on this task.