Lesson 17 · 24 in the route · 5 min read · Level 2

Mac vs Windows

The same tasks on both machines; cloud vs local in claude -p; a local coding model vs Claude Code.

Goal

Know with numbers what is worth sending to the Mac: tests, builds, git, a real claude -p job, and whether a local 9B model replaces claude -p for code.

Why

“The Mac is faster” is no basis for a decision. You have to separate two things the clock mixes together: the cloud model time (the same on any machine) and the local time (starting Node, hooks, MCP servers, spawning processes). Only the second depends on the machine.

Machines: Mac mini M2 Pro 16 GB · Windows 10, Intel i5-8600, 48 GB, Git Bash.

Steps

1. Same code on both: `git bundle`

PS> git -C ..\orquestador bundle create <bundles>\orquestador.bundle --all
PS> git -C ..\enjambre    bundle create <bundles>\enjambre.bundle --all
PS> scp <bundles>\*.bundle mac:~/modelos/benchcmp/
PS> python scripts\bench_compare.py <bundles> <work> windows --claude
mac$ python scripts/bench_compare.py ~/modelos/benchcmp ~/modelos/benchcmp/work mac --claude

Both sides clone from the same bundle: same commits (orquestador 89ae2fd, enjambre 6a14b6e), with no dependence on GitHub or the keychain.

2. Local tools (median of 3)

Task	Windows	Mac	Mac faster by
orquestador tests: statusline	15.40 s	1.08 s	14×
orquestador tests: hooks	17.59 s	1.03 s	17×
enjambre tests (Python, 38)	0.55 s	0.35 s	1.6×
`build-kit.sh` build	16.98 s	0.62 s	27×
local git clone	0.24 s	0.03 s	8×
git status ×20	1.42 s	0.14 s	10×
git gc —aggressive	1.88 s	0.85 s	2.2×

The huge gap is in the bash scripts; in plain Python it is barely 1.6×. A hypothesis consistent with that (not measured separately): on Windows every jq, sed or date goes through MSYS’s fork emulation, which is expensive, and these tests launch hundreds of processes.

3. `claude -p`: the cloud on one side, local on the other

The JSON from claude -p --output-format json includes duration_api_ms. Local = clock − API:

(capture of the real setup: not published; see docs/CAPTURAS.md for the demo-mode ones)

	Clock	Cloud	Local	Cost
Windows	45.4 s	15.0 s	30.4 s	US$ 0.28
Mac	15.1 s	11.1 s	4.0 s	US$ 0.18

Is it the machine or the configuration? Windows has 13 MCP servers and the Mac 8. With a trivial prompt, 2 runs per cell:

Local startup	Windows	Mac
Normal	18.2 / 18.7 s	4.1 / 3.6 s
No MCP (`--strict-mcp-config`)	13.6 / 12.9 s	1.8 / 1.6 s

The MCPs add ~5 s on Windows; the rest (~13 s against ~1.7 s) is the platform.

4. Local code model vs `claude -p`

Three functions with hidden tests that neither side sees: amounts in soles, merging intervals, RUC check digit. The local model gets one attempt and one retry with the test output; claude -p is an agent that can write and run its own tests.

PS> python scripts/bench_code.py            # local sin razonamiento + claude -p
PS> python scripts/bench_code.py --think --skip-claude

Engine	Passed	Time	Cost
`claude -p` (Opus 5.5)	3/3	132 s	US$ 0.91
Local Qwen3.5-9B	1/3	155 s	US$ 0
Local Qwen3.5-9B with thinking	1/3	751 s	US$ 0

No sugarcoating: for code with exact rules, the local 9B does not replace claude -p. It got merge_intervals right on the first try; it failed the thousands grouping and the check digit even with a retry. Reasoning more made it 5× slower without getting more right.

What actually went wrong

build-kit.sh does not run on the Mac: BSD sed rejects line 55 (RE error: parentheses not balanced, a \) with no \(). It was measured with GNU sed (brew install gnu-sed); the bug was noted to be fixed in orquestador.
“2 failures on Windows” that were not Windows’s fault: two statusline tests look up the session by directory and the test clone lived at an odd temporary path. On the normal checkout of the same commit: 50/0.
claude -p from Python on Windows received half the prompt: claude is a .cmd and cmd.exe cuts the arguments at the first line break. Fix: the prompt goes through stdin.
Deleting the test clone failed on Windows: git packs are read-only. Fix: remove the attribute and retry inside rmtree.

How to verify

PS> python scripts\bench_compare.py <bundles> <work> windows --claude
mac$ python scripts/bench_compare.py ~/modelos/benchcmp ~/modelos/benchcmp/work mac --claude

Expected (medians of 3, step 2): statusline tests 15.40 s on Windows against 1.08 s on the Mac; build-kit.sh 16.98 s against 0.62 s; enjambre tests (Python) 0.55 s against 0.35 s. claude -p with the real job: local 30.4 s on Windows against 4.0 s on the Mac. Both sides clone from the same bundle (orquestador 89ae2fd, enjambre 6a14b6e). With a small n, the numbers are indicative only.

Limits

Small n (1 run of the real job per machine). Outside load: Photos indexing on the Mac, a browser and another Claude session on Windows. The times are a ceiling, not an average.

Self-check

Why is duration_api_ms subtracted from the clock and not the other way round?
If you remove all the MCPs, how much of the Windows–Mac gap is left? What does that tell you?
What kind of task would you give the local model today, and which one never?

Saved only in this browser.