Benchmarks · Measured July 2026 · 6 min
We raced five ways to find one file. Four are still running.
One file. One 4.47-million-file Windows machine. One stopwatch. We handed the same “where is it?” to the five tools a developer actually reaches for (grep's faster cousin ripgrep, Windows Search, VS Code, Claude Code, and a pre-built index) and started the clock. The gap between last place and first is not a percentage. It is the difference between a coffee break and a heartbeat.
The race, to scale
Every bar is real seconds. The winner is a dot.
No log scale, no tricks. The axis below is plain wall-clock time, zero to one hundred seconds. At that scale the pre-built index lands as a single point on the start line, ripgrep (the fastest file-walker ever shipped) nearly runs out of room, and the smartest contestant leaves the page entirely.
Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4.47M files. One measured lookup per tool. Claude Code's bar is drawn to the page edge; it actually ends at 193.5 s.
The finishing order
Same file. Five very different clocks.
“Find one file” is an honest question with dishonest edges: some tools look by name, some look by what is inside. So we let each play to its strength and timed whatever it did. VS Code matched the name. Windows Search used its own system index. ripgrep and Claude Code went looking through the contents. Every number below is measured on the same drive.
| Finisher | How it looks for the file | Time | vs. the index |
|---|---|---|---|
| Interlinked | asks the pre-built index, no walk | 85 μs | 1× |
| VS Code · Ctrl+P | fuzzy filename over the open workspace | 35 s | 412,000× |
| Windows Search | the system index · median of 20 queries | 67 s | ~480,000× |
| ripgrep | reads the contents of every file, in parallel | 93.8 s | 58,625× † |
| Claude Code · Opus | loops the same shell walkers, 71 calls | 3 m 13 s | 2,276,000× |
† ripgrep's 58,625× is the average across our five-query content set, where the index answers in ~8 ms (7 to 9 ms typical). The single query shown (93.8 s on ripgrep) returns from the index in 16 ms. Filename lookups: 85 μs for a single file (2 μs best case), 139 μs median across 20 queries. Full per-tool methodology in the 412,000× VS Code write-up and the ripgrep comparison.
The two names you already know
The best walker and the smartest agent finished last.
ripgrep is the crown of walking the disk: it compares many bytes at once, spreads across every core, maps files into memory instead of reading them in chunks. It is genuinely the fastest way ever built to open, scan, and close files one by one. And on 4.47 million of them it still needs 93.8 seconds, because opening and reading every file is the floor, and that floor is measured in minutes.
Claude Code is the smartest thing in the race, and it came in last. Ask it to find a file and it does what a careful human would: it reaches for the shell tools it has (find, glob, grep), fires one, reads the output, decides what to try next, and repeats. Seventy-one times. Intelligence did not beat the walk. It just orchestrated more of it.
Interlinked
one lookup · no agent involved
Claude Code · Opus
71 tool calls · 2,276,000× slower
The proof line
Same file, same machine, same content query. Hand that same lookup to the index and it comes back before the walk would have finished listing the first directory.
Why first place is a dot
The walk already happened, once, when nobody was waiting.
The four tools that are still running all do the same thing the moment you press enter: they start walking the disk from cold. The index did that walk ahead of time and kept the answer warm in a background service. Your query does not traverse anything. It reads a finished answer. That is the entire difference between 85 microseconds and three minutes: the work was already done.
It stays honest as files change. Save a file and it is searchable in about a millisecond (under 30 milliseconds in the worst case), so the warm answer is never stale. And keeping it ready is cheap: the whole engine sits around 44 MB at rest, less than one browser tab, and the index on disk stays under 1% of the drive (about 24 GB on a 4 TB machine).
The payoff
Make finding free and the agent stops flailing.
Same model, same machine: one thing changed, the floor it stands on. Over a real session, that Claude Code spent 6 minutes 57 seconds and 71 tool calls hunting for files, roughly 58% of its tokens spent not on thinking but on looking. Give the same agent the index over its tool connection and that collapses to one call, and the bill for finding things falls to about nothing.
On a three-minute-per-lookup budget an agent asks two or three careful questions, then stops. On an 85-microsecond budget it can check every callsite, verify every import, ask speculative questions and throw away the answers that did not help, all before a walker finishes its first pass. The model did not get smarter. The floor did.
CHEAP QUESTIONS ARE A DIFFERENT WORKFLOW THAN EXPENSIVE ONES.
Do this today
Give every agent the short clock.
Install once. One signed installer wires up 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing, no keys.
Let it index. Filename search works immediately; contents fill in behind it. At rest the whole engine sits around 44 MB, less than a browser tab.
Ask freely. Whole-machine when you don't know where something lives; scoped to a folder when you do. Either way the answer comes back in the microsecond range.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.