The benchmark renders the exact same interactive stats-01-animated registry
block with React 19.2.7 and Tailwind 4.3.2 in four production builds. It measures
delivery overhead for one real component graph; it is not a universal framework
ranking.
| Framework | Build ms | Client gzip | JS transferred | Hydration ready | LCP | TTFB |
|---|---|---|---|---|---|---|
| Next.js | 5,436.2 | 184.6 KiB | 143.5 KiB | 1,254.3 ms | 468 ms | 2.3 ms |
| Vite | 621.4 | 73.2 KiB | 69.6 KiB | 822.4 ms | 832 ms | 1.3 ms |
| Astro | 1,586.2 | 74.6 KiB | 71.5 KiB | 981.9 ms | 400 ms | 1.5 ms |
| TanStack Start | 1,208.2 | 112 KiB | 108.7 KiB | 1,043.3 ms | 424 ms | 3.2 ms |
Median of 5 runs. Generated 29 Aug 2026, 10:33 UTC — Apple M2(8 logical cores). Lower is better. Exact versions: Next.js 16.3.3, Vite 8.2.2, Astro 7.2.9, TanStack Start 1.168.49.
Profile
Each framework gets a clean production build and five fresh Chromium contexts. Chrome DevTools Protocol applies a 4× CPU slowdown, 150 ms round-trip latency, 195.3 KiB/s download and 91.6 KiB/s upload. The table reports medians.
- Client gzip is the sum of production client JS and CSS compressed with gzip. Server output is intentionally excluded.
- JS transferred is the same-origin JavaScript reported by the browser Resource Timing API with cache disabled.
- Hydration ready is
performance.now()at the shared scenario’s first React effect. In the Vite SPA this includes the initial client render; in the three SSR/static-HTML variants it marks hydration. - LCP and TTFB come from browser Performance APIs under the same throttled profile.
Reproduce it
The isolated workspace lives in benchmarks/frameworks and exact-pins every
framework. From the repository root:
cd benchmarks/frameworks
pnpm install --frozen-lockfile
pnpm exec playwright install chromium
pnpm benchmarkBefore each run the harness copies the current block and its transitive registry
source into an ignored generated directory, so all four fixtures compile the
same product code without drifting copies. It writes the raw profile, host,
versions, samples summary and results to results/latest.json.
Compare changes only when runs use the same host and profile. These local preview servers do not measure CDN caching, serverless cold starts, production network topology or developer experience.