Every tool checks every file of each corpus, using all cores, on an M1 Max. Times are the mean of both run orders. Embedded code is excluded, so that every tool does the same amount of work.
| Corpus | Files | tidy | bun fmt | oxfmt | dprint | Prettier |
|---|---|---|---|---|---|---|
| JavaScript | 3,906 | 78 ms | 100 ms | 123 ms | 8,528 ms | |
| TypeScript | 3,309 | 58 ms | 85 ms | 96 ms | 7,523 ms | |
| JSON | 4,569 | 86 ms | 149 ms | 114 ms | 323 ms | 3,528 ms |
| CSS | 300 | 7.2 ms | 35 ms | 47 ms | 25 ms | 654 ms |
| SCSS | 101 | 4.6 ms | 32 ms | 46 ms | 11 ms | 340 ms |
| Less | 396 | 9.7 ms | 37 ms | 48 ms | 58 ms | 1,491 ms |
| HTML | 151 | 6.4 ms | 34 ms | 237 ms | 14 ms | 327 ms |
| Vue | 230 | 7.3 ms | 34 ms | 435 ms | 205 ms | 684 ms |
| Handlebars | 57 | 3.7 ms | 30 ms | 201 ms | 279 ms | |
| Markdown | 2,834 | 37 ms | 71 ms | 80 ms | 188 ms | 6,618 ms |
| YAML | 435 | 24 ms | 35 ms | 46 ms | 104 ms | 1,271 ms |
| TOML | 375 | 8.6 ms | 36 ms | 49 ms | 22 ms | |
| GraphQL | 2,511 | 37 ms | 63 ms | 72 ms | 271 ms | 1,988 ms |
| XML | 6,784 | 130 ms | 616 ms | 6,017 ms | ||
| Dockerfiles | 25 | 3.6 ms | 3.7 ms | 458 ms | ||
| INI | 75 | 3.7 ms | 147 ms | |||
| All but XML | 12,071 | 280 ms | 569 ms | 1,232 ms | 3,474 ms | 18,548 ms |
A blank cell means the tool does not format that corpus.
With embedded code
Formatting the code inside each file as well — styles and scripts in HTML, code blocks in Markdown — widens the gap, because that is more parsing for everyone and because not every tool does it:
| Tool | All but XML |
|---|---|
| tidy | 286 ms |
| bun fmt | 609 ms |
| oxfmt | 2.4 s |
| Prettier | 30.9 s |
tidy adds 6 ms to do it. Prettier adds 12 seconds.
Reading these numbers
XML is listed apart because its corpus is 6,784 files, over half the total, which would dominate a combined figure and flatter whichever tool is best at XML.
The ranking depends on what you format. bun fmt and oxfmt are fast on JavaScript and slow on markup: oxfmt takes 237 ms on 151 HTML files, where tidy takes 6.4 ms. A JavaScript-only repository sees a much smaller gap than these totals suggest.
Throughput, not latency. Every number is a whole corpus with every core busy. Formatting one file on save is a different measurement, dominated by process start — which is where a native binary is furthest ahead of a Node process.
Method
- All cores, unless the full results say otherwise. The repository also has one-thread numbers.
- Each pair of tools is run in both orders and the two means averaged, so that a warm page cache does not favour whichever ran second.
- Every tool runs in check mode where it has one, so none of them pays for writing files.
- The corpora are real files from open-source projects, not generated input.
The full results
Tables with one-thread times, and with Biome, taplo, yamlfmt, lightningcss and oxc's CSS formatter, are in bench/results/format.
The scripts are bench/run.ts and bench/js.ts. Run them yourself:
$ bun bench/run.tsVersions
| Tool | Version |
|---|---|
| Prettier | 3.9.9 |
| bun fmt | 1.4 |
| oxfmt | 0.75.0 |
| dprint | with its TOML, Dockerfile, Markdown and JSON plugins |
Numbers move as these tools change. If a result here is stale, the scripts above are the way to check.