diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml
new file mode 100644
index 0000000..cae6a1f
--- /dev/null
+++ b/.github/workflows/bench.yml
@@ -0,0 +1,56 @@
+name: Benchmarks
+
+on:
+  # Run on pushes to main (to keep report up to date)
+  push:
+    branches: [main]
+  # Run weekly on Mondays at 06:00 UTC
+  schedule:
+    - cron: "0 6 * * 1"
+  # Allow manual trigger
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: write
+
+jobs:
+  bench:
+    name: Run Benchmarks
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Install dependencies
+        run: bun install --frozen-lockfile
+
+      - name: Run benchmarks and generate report
+        run: bun run bench:report
+
+      - name: Upload JSON results
+        uses: actions/upload-artifact@v4
+        with:
+          name: bench-results
+          path: reports/bench-results.json
+          retention-days: 90
+
+      - name: Commit updated report
+        run: |
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+
+          git add reports/benchmarks.md
+
+          if git diff --staged --quiet; then
+            echo "No changes to benchmark report"
+          else
+            git commit -m "docs: update benchmark report"
+            git push
+          fi
diff --git a/.gitignore b/.gitignore
index 0d003c2..fd26d33 100644
--- a/.gitignore
+++ b/.gitignore
@@ -52,5 +52,8 @@ debug/
 fixtures/benchmarks/
 fixtures/private/
 
+# Benchmark JSON results (machine-specific)
+reports/bench-results.json
+
 # Temporary files
 tmp/
diff --git a/benchmarks/comparison.bench.ts b/benchmarks/comparison.bench.ts
index 255a57d..355c33d 100644
--- a/benchmarks/comparison.bench.ts
+++ b/benchmarks/comparison.bench.ts
@@ -9,10 +9,12 @@ import { PDFDocument } from "pdf-lib";
 import { bench, describe } from "vitest";
 
 import { PDF } from "../src";
-import { loadFixture, getHeavyPdf } from "./fixtures";
+import { getHeavyPdf, getSynthetic100, getSynthetic2000, loadFixture } from "./fixtures";
 
-// Pre-load fixture
+// Pre-load fixtures
 const pdfBytes = await getHeavyPdf();
+const synthetic100 = await getSynthetic100();
+const synthetic2000 = await getSynthetic2000();
 
 describe("Load PDF", () => {
   bench("libpdf", async () => {
@@ -119,3 +121,145 @@ describe("Load, modify, and save PDF", () => {
     await pdf.save();
   });
 });
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Page splitting comparison (issue #26)
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Extract single page from 100-page PDF", () => {
+  bench("libpdf", async () => {
+    const pdf = await PDF.load(synthetic100);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+
+  bench("pdf-lib", async () => {
+    const pdf = await PDFDocument.load(synthetic100);
+    const newDoc = await PDFDocument.create();
+    const [page] = await newDoc.copyPages(pdf, [0]);
+    newDoc.addPage(page);
+    await newDoc.save();
+  });
+});
+
+describe("Split 100-page PDF into single-page PDFs", () => {
+  bench(
+    "libpdf",
+    async () => {
+      const pdf = await PDF.load(synthetic100);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    "pdf-lib",
+    async () => {
+      const pdf = await PDFDocument.load(synthetic100);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const newDoc = await PDFDocument.create();
+        const [page] = await newDoc.copyPages(pdf, [i]);
+        newDoc.addPage(page);
+        await newDoc.save();
+      }
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
+
+describe(`Split 2000-page PDF into single-page PDFs (${(synthetic2000.length / 1024 / 1024).toFixed(1)}MB)`, () => {
+  bench(
+    "libpdf",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 0, iterations: 1, time: 0 },
+  );
+
+  bench(
+    "pdf-lib",
+    async () => {
+      const pdf = await PDFDocument.load(synthetic2000);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const newDoc = await PDFDocument.create();
+        const [page] = await newDoc.copyPages(pdf, [i]);
+        newDoc.addPage(page);
+        await newDoc.save();
+      }
+    },
+    { warmupIterations: 0, iterations: 1, time: 0 },
+  );
+});
+
+describe("Copy 10 pages between documents", () => {
+  bench("libpdf", async () => {
+    const source = await PDF.load(synthetic100);
+    const dest = PDF.create();
+    const indices = Array.from({ length: 10 }, (_, i) => i);
+    await dest.copyPagesFrom(source, indices);
+    await dest.save();
+  });
+
+  bench("pdf-lib", async () => {
+    const source = await PDFDocument.load(synthetic100);
+    const dest = await PDFDocument.create();
+    const indices = Array.from({ length: 10 }, (_, i) => i);
+    const pages = await dest.copyPages(source, indices);
+
+    for (const page of pages) {
+      dest.addPage(page);
+    }
+
+    await dest.save();
+  });
+});
+
+describe("Merge 2 x 100-page PDFs", () => {
+  bench(
+    "libpdf",
+    async () => {
+      const merged = await PDF.merge([synthetic100, synthetic100]);
+      await merged.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    "pdf-lib",
+    async () => {
+      const doc1 = await PDFDocument.load(synthetic100);
+      const doc2 = await PDFDocument.load(synthetic100);
+      const merged = await PDFDocument.create();
+
+      const pages1 = await merged.copyPages(doc1, doc1.getPageIndices());
+
+      for (const page of pages1) {
+        merged.addPage(page);
+      }
+
+      const pages2 = await merged.copyPages(doc2, doc2.getPageIndices());
+
+      for (const page of pages2) {
+        merged.addPage(page);
+      }
+
+      await merged.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
diff --git a/benchmarks/copying.bench.ts b/benchmarks/copying.bench.ts
new file mode 100644
index 0000000..034b6c2
--- /dev/null
+++ b/benchmarks/copying.bench.ts
@@ -0,0 +1,94 @@
+/**
+ * PDF page-copying and merging benchmarks.
+ *
+ * Tests the performance of copying pages between documents and merging
+ * multiple PDFs. These operations are closely related to splitting
+ * (issue #26) and represent the other side of the workflow.
+ */
+
+import { bench, describe } from "vitest";
+
+import { PDF } from "../src";
+import { getSynthetic100, loadFixture, mediumPdfPath } from "./fixtures";
+
+// Pre-load fixtures
+const mediumPdf = await loadFixture(mediumPdfPath);
+const synthetic100 = await getSynthetic100();
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Page copying
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Copy pages between documents", () => {
+  bench("copy 1 page", async () => {
+    const source = await PDF.load(mediumPdf);
+    const dest = PDF.create();
+    await dest.copyPagesFrom(source, [0]);
+    await dest.save();
+  });
+
+  bench("copy 10 pages from 100-page PDF", async () => {
+    const source = await PDF.load(synthetic100);
+    const dest = PDF.create();
+    const indices = Array.from({ length: 10 }, (_, i) => i);
+    await dest.copyPagesFrom(source, indices);
+    await dest.save();
+  });
+
+  bench(
+    "copy all 100 pages",
+    async () => {
+      const source = await PDF.load(synthetic100);
+      const dest = PDF.create();
+      const indices = Array.from({ length: 100 }, (_, i) => i);
+      await dest.copyPagesFrom(source, indices);
+      await dest.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Self-copy (page duplication)
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Duplicate pages within same document", () => {
+  bench("duplicate page 0", async () => {
+    const pdf = await PDF.load(mediumPdf);
+    await pdf.copyPagesFrom(pdf, [0]);
+    await pdf.save();
+  });
+
+  bench("duplicate all pages (double the document)", async () => {
+    const pdf = await PDF.load(mediumPdf);
+    const indices = Array.from({ length: pdf.getPageCount() }, (_, i) => i);
+    await pdf.copyPagesFrom(pdf, indices);
+    await pdf.save();
+  });
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Merging
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Merge PDFs", () => {
+  bench("merge 2 small PDFs", async () => {
+    const merged = await PDF.merge([mediumPdf, mediumPdf]);
+    await merged.save();
+  });
+
+  bench("merge 10 small PDFs", async () => {
+    const sources = Array.from({ length: 10 }, () => mediumPdf);
+    const merged = await PDF.merge(sources);
+    await merged.save();
+  });
+
+  bench(
+    "merge 2 x 100-page PDFs",
+    async () => {
+      const merged = await PDF.merge([synthetic100, synthetic100]);
+      await merged.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
diff --git a/benchmarks/fixtures.ts b/benchmarks/fixtures.ts
index b18e21c..4cf11f2 100644
--- a/benchmarks/fixtures.ts
+++ b/benchmarks/fixtures.ts
@@ -1,12 +1,16 @@
 /**
  * Benchmark fixture helpers.
  *
- * Provides utilities for loading PDF fixtures for benchmarks.
+ * Provides utilities for loading PDF fixtures for benchmarks,
+ * including synthetic large PDFs built by copying pages from
+ * existing fixtures.
  */
 
 import { existsSync, mkdirSync, writeFileSync } from "node:fs";
 import { readFile } from "node:fs/promises";
 
+import { PDF } from "../src";
+
 // Heavy PDF - downloaded on first run (~10MB)
 const HEAVY_PDF_PATH = "fixtures/benchmarks/cc-journalists-guide.pdf";
 const HEAVY_PDF_URL =
@@ -15,6 +19,11 @@ const HEAVY_PDF_URL =
 // Fallback large PDF - use existing fixture from pdfbox malformed tests (2MB)
 const LARGE_PDF_FALLBACK = "fixtures/malformed/pdfbox/PDFBOX-3947.pdf";
 
+// Synthetic PDFs - generated on first run, cached locally
+const SYNTHETIC_DIR = "fixtures/benchmarks";
+const SYNTHETIC_100_PATH = `${SYNTHETIC_DIR}/synthetic-100p.pdf`;
+const SYNTHETIC_2000_PATH = `${SYNTHETIC_DIR}/synthetic-2000p.pdf`;
+
 /**
  * Load a fixture file as bytes.
  */
@@ -65,6 +74,90 @@ export async function getLargePdf(): Promise<Uint8Array> {
   return loadFixture(LARGE_PDF_FALLBACK);
 }
 
+/**
+ * Build a synthetic PDF with the given number of pages by copying
+ * pages from sample.pdf. Each page gets unique text to simulate
+ * real-world content variation.
+ */
+async function buildSyntheticPdf(pageCount: number): Promise<Uint8Array> {
+  const sourceBytes = await loadFixture(mediumPdfPath);
+  const source = await PDF.load(sourceBytes);
+  const sourcePageCount = source.getPageCount();
+
+  // Start by copying the source pages
+  const pdf = await PDF.load(sourceBytes);
+
+  // Copy pages from source repeatedly until we reach the target count
+  const pagesNeeded = pageCount - sourcePageCount;
+
+  if (pagesNeeded > 0) {
+    // Build an array of source page indices to copy in bulk
+    const indices: number[] = [];
+
+    for (let i = 0; i < pagesNeeded; i++) {
+      indices.push(i % sourcePageCount);
+    }
+
+    await pdf.copyPagesFrom(source, indices);
+  }
+
+  // Add unique text to each page so content varies
+  for (let i = 0; i < pdf.getPageCount(); i++) {
+    const page = pdf.getPage(i);
+
+    if (page) {
+      page.drawText(`Page ${i + 1} of ${pageCount}`, {
+        x: 50,
+        y: 20,
+        font: "Helvetica",
+        size: 8,
+      });
+    }
+  }
+
+  return pdf.save();
+}
+
+/**
+ * Get or create a synthetic PDF cached to disk.
+ */
+async function getOrCreateSynthetic(path: string, pageCount: number): Promise<Uint8Array> {
+  if (existsSync(path)) {
+    return loadFixture(path);
+  }
+
+  console.log(`Building synthetic ${pageCount}-page PDF...`);
+  const start = performance.now();
+
+  const bytes = await buildSyntheticPdf(pageCount);
+
+  mkdirSync(SYNTHETIC_DIR, { recursive: true });
+  writeFileSync(path, bytes);
+
+  const elapsed = ((performance.now() - start) / 1000).toFixed(1);
+  const size = (bytes.length / 1024 / 1024).toFixed(1);
+
+  console.log(`Cached ${pageCount}-page PDF to ${path} (${size}MB) in ${elapsed}s`);
+
+  return bytes;
+}
+
+/**
+ * Get a synthetic 100-page PDF.
+ * Built by copying pages from sample.pdf. Cached on disk after first build.
+ */
+export async function getSynthetic100(): Promise<Uint8Array> {
+  return getOrCreateSynthetic(SYNTHETIC_100_PATH, 100);
+}
+
+/**
+ * Get a synthetic 2000-page PDF.
+ * Built by copying pages from sample.pdf. Cached on disk after first build.
+ */
+export async function getSynthetic2000(): Promise<Uint8Array> {
+  return getOrCreateSynthetic(SYNTHETIC_2000_PATH, 2000);
+}
+
 // Pre-load common fixtures
 export const smallPdfPath = "fixtures/basic/rot0.pdf";
 export const mediumPdfPath = "fixtures/basic/sample.pdf";
diff --git a/benchmarks/splitting.bench.ts b/benchmarks/splitting.bench.ts
new file mode 100644
index 0000000..23f5708
--- /dev/null
+++ b/benchmarks/splitting.bench.ts
@@ -0,0 +1,119 @@
+/**
+ * PDF page-splitting benchmarks.
+ *
+ * Tests the performance of splitting a PDF into individual single-page PDFs.
+ * This is the primary benchmark requested in issue #26 for users who need
+ * to split 2000+ page documents at high throughput.
+ *
+ * Scenarios:
+ * - Extract single page (baseline)
+ * - Split 100-page PDF into individual pages
+ * - Split 2000-page PDF into individual pages
+ */
+
+import { bench, describe } from "vitest";
+
+import { PDF } from "../src";
+import { getSynthetic100, getSynthetic2000, loadFixture, mediumPdfPath } from "./fixtures";
+
+// Pre-load fixtures outside benchmarks to isolate I/O from measurements
+const mediumPdf = await loadFixture(mediumPdfPath);
+const synthetic100 = await getSynthetic100();
+const synthetic2000 = await getSynthetic2000();
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Single page extraction (baseline)
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Extract single page", () => {
+  bench("extractPages (1 page from small PDF)", async () => {
+    const pdf = await PDF.load(mediumPdf);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+
+  bench("extractPages (1 page from 100-page PDF)", async () => {
+    const pdf = await PDF.load(synthetic100);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+
+  bench("extractPages (1 page from 2000-page PDF)", async () => {
+    const pdf = await PDF.load(synthetic2000);
+    const extracted = await pdf.extractPages([0]);
+    await extracted.save();
+  });
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Full split: every page into its own PDF
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Split into single-page PDFs", () => {
+  bench(
+    `split 100-page PDF (${(synthetic100.length / 1024 / 1024).toFixed(1)}MB)`,
+    async () => {
+      const pdf = await PDF.load(synthetic100);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    `split 2000-page PDF (${(synthetic2000.length / 1024 / 1024).toFixed(1)}MB)`,
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const pageCount = pdf.getPageCount();
+
+      for (let i = 0; i < pageCount; i++) {
+        const single = await pdf.extractPages([i]);
+        await single.save();
+      }
+    },
+    { warmupIterations: 0, iterations: 1, time: 0 },
+  );
+});
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Batch extraction: extract ranges of pages
+// ─────────────────────────────────────────────────────────────────────────────
+
+describe("Batch page extraction", () => {
+  bench(
+    "extract first 10 pages from 2000-page PDF",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const indices = Array.from({ length: 10 }, (_, i) => i);
+      const extracted = await pdf.extractPages(indices);
+      await extracted.save();
+    },
+    { warmupIterations: 1, iterations: 5 },
+  );
+
+  bench(
+    "extract first 100 pages from 2000-page PDF",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const indices = Array.from({ length: 100 }, (_, i) => i);
+      const extracted = await pdf.extractPages(indices);
+      await extracted.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+
+  bench(
+    "extract every 10th page from 2000-page PDF (200 pages)",
+    async () => {
+      const pdf = await PDF.load(synthetic2000);
+      const indices = Array.from({ length: 200 }, (_, i) => i * 10);
+      const extracted = await pdf.extractPages(indices);
+      await extracted.save();
+    },
+    { warmupIterations: 1, iterations: 3 },
+  );
+});
diff --git a/package.json b/package.json
index b511e67..230f00b 100644
--- a/package.json
+++ b/package.json
@@ -46,6 +46,7 @@
   },
   "scripts": {
     "bench": "vitest bench",
+    "bench:report": "bun run scripts/bench-report.ts",
     "build": "tsdown",
     "docs:build": "bun run --cwd apps/docs build",
     "docs:dev": "bun run --cwd apps/docs dev",
diff --git a/reports/benchmarks.md b/reports/benchmarks.md
new file mode 100644
index 0000000..b0b0010
--- /dev/null
+++ b/reports/benchmarks.md
@@ -0,0 +1,235 @@
+# Benchmark Report
+
+> Generated on 2026-02-16 at 12:50:00 UTC
+>
+> System: darwin | Apple M4 Pro (12 cores) | 24GB RAM | Bun 1.3.5
+
+---
+
+## Contents
+
+- [Comparison](#comparison)
+- [Copying](#copying)
+- [Drawing](#drawing)
+- [Forms](#forms)
+- [Loading](#loading)
+- [Saving](#saving)
+- [Splitting](#splitting)
+
+## Comparison
+
+### Load PDF
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |   895.0 |  1.12ms |  1.59ms | ±1.07% |     448 |
+| pdf-lib   |    36.7 | 27.21ms | 29.03ms | ±1.90% |      19 |
+
+- **libpdf** is 24.35x faster than pdf-lib
+
+### Create blank PDF
+
+| Benchmark | ops/sec | Mean |   p99 |    RME | Samples |
+| :-------- | ------: | ---: | ----: | -----: | ------: |
+| libpdf    |   38.6K | 26us |  45us | ±0.84% |  19,283 |
+| pdf-lib   |   10.3K | 97us | 461us | ±1.82% |   5,172 |
+
+- **libpdf** is 3.73x faster than pdf-lib
+
+### Add 10 pages
+
+| Benchmark | ops/sec |  Mean |   p99 |    RME | Samples |
+| :-------- | ------: | ----: | ----: | -----: | ------: |
+| libpdf    |   19.1K |  52us |  87us | ±0.91% |   9,562 |
+| pdf-lib   |    6.3K | 158us | 770us | ±2.55% |   3,173 |
+
+- **libpdf** is 3.01x faster than pdf-lib
+
+### Draw 50 rectangles
+
+| Benchmark | ops/sec |   Mean |    p99 |    RME | Samples |
+| :-------- | ------: | -----: | -----: | -----: | ------: |
+| pdf-lib   |    2.2K |  458us | 1.80ms | ±3.80% |   1,093 |
+| libpdf    |   627.4 | 1.59ms | 2.30ms | ±1.44% |     314 |
+
+- **pdf-lib** is 3.48x faster than libpdf
+
+### Load and save PDF
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |   909.3 |  1.10ms |  1.48ms | ±0.88% |     456 |
+| pdf-lib   |    22.0 | 45.45ms | 58.21ms | ±6.77% |      11 |
+
+- **libpdf** is 41.33x faster than pdf-lib
+
+### Load, modify, and save PDF
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |    37.9 | 26.38ms | 33.50ms | ±5.76% |      20 |
+| pdf-lib   |    23.1 | 43.25ms | 44.89ms | ±1.52% |      12 |
+
+- **libpdf** is 1.64x faster than pdf-lib
+
+### Extract single page from 100-page PDF
+
+| Benchmark | ops/sec |   Mean |    p99 |    RME | Samples |
+| :-------- | ------: | -----: | -----: | -----: | ------: |
+| libpdf    |   503.4 | 1.99ms | 3.10ms | ±1.55% |     252 |
+| pdf-lib   |   155.9 | 6.41ms | 7.44ms | ±1.53% |      79 |
+
+- **libpdf** is 3.23x faster than pdf-lib
+
+### Split 100-page PDF into single-page PDFs
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |    35.3 | 28.29ms | 29.79ms | ±1.35% |      18 |
+| pdf-lib   |    35.0 | 28.58ms | 32.30ms | ±3.07% |      18 |
+
+- **libpdf** is 1.01x faster than pdf-lib
+
+### Copy 10 pages between documents
+
+| Benchmark | ops/sec |   Mean |     p99 |    RME | Samples |
+| :-------- | ------: | -----: | ------: | -----: | ------: |
+| libpdf    |   334.0 | 2.99ms |  3.55ms | ±1.12% |     168 |
+| pdf-lib   |   103.7 | 9.64ms | 14.95ms | ±3.70% |      52 |
+
+- **libpdf** is 3.22x faster than pdf-lib
+
+### Merge 2 x 100-page PDFs
+
+| Benchmark | ops/sec |    Mean |     p99 |    RME | Samples |
+| :-------- | ------: | ------: | ------: | -----: | ------: |
+| libpdf    |    47.1 | 21.23ms | 24.31ms | ±2.08% |      24 |
+| pdf-lib   |    22.9 | 43.64ms | 47.48ms | ±2.23% |      12 |
+
+- **libpdf** is 2.06x faster than pdf-lib
+
+## Copying
+
+### Copy pages between documents
+
+| Benchmark                       | ops/sec |    Mean |     p99 |    RME | Samples |
+| :------------------------------ | ------: | ------: | ------: | -----: | ------: |
+| copy 1 page                     |    2.3K |   429us |   735us | ±1.26% |   1,166 |
+| copy 10 pages from 100-page PDF |   344.1 |  2.91ms |  3.57ms | ±1.12% |     173 |
+| copy all 100 pages              |    92.3 | 10.84ms | 13.86ms | ±1.96% |      47 |
+
+- **copy 1 page** is 6.78x faster than copy 10 pages from 100-page PDF
+- **copy 1 page** is 25.27x faster than copy all 100 pages
+
+### Duplicate pages within same document
+
+| Benchmark                                 | ops/sec |  Mean |   p99 |    RME | Samples |
+| :---------------------------------------- | ------: | ----: | ----: | -----: | ------: |
+| duplicate all pages (double the document) |    2.2K | 461us | 798us | ±0.89% |   1,086 |
+| duplicate page 0                          |    2.2K | 464us | 758us | ±0.77% |   1,078 |
+
+- **duplicate all pages (double the document)** is 1.01x faster than duplicate page 0
+
+### Merge PDFs
+
+| Benchmark               | ops/sec |    Mean |     p99 |    RME | Samples |
+| :---------------------- | ------: | ------: | ------: | -----: | ------: |
+| merge 2 small PDFs      |    1.4K |   712us |  1.12ms | ±1.00% |     702 |
+| merge 10 small PDFs     |   254.5 |  3.93ms |  5.73ms | ±2.08% |     128 |
+| merge 2 x 100-page PDFs |    48.6 | 20.58ms | 26.32ms | ±3.33% |      25 |
+
+- **merge 2 small PDFs** is 5.51x faster than merge 10 small PDFs
+- **merge 2 small PDFs** is 28.89x faster than merge 2 x 100-page PDFs
+
+## Drawing
+
+| Benchmark                           | ops/sec |   Mean |    p99 |    RME | Samples |
+| :---------------------------------- | ------: | -----: | -----: | -----: | ------: |
+| draw 100 lines                      |   399.8 | 2.50ms | 2.74ms | ±0.51% |     200 |
+| draw 100 rectangles                 |   360.3 | 2.78ms | 3.55ms | ±1.22% |     181 |
+| draw 100 circles                    |   279.7 | 3.58ms | 4.42ms | ±1.33% |     140 |
+| draw 100 text lines (standard font) |   259.5 | 3.85ms | 4.29ms | ±0.61% |     130 |
+| create 10 pages with mixed content  |   193.0 | 5.18ms | 6.40ms | ±1.35% |      97 |
+
+- **draw 100 lines** is 1.11x faster than draw 100 rectangles
+- **draw 100 lines** is 1.43x faster than draw 100 circles
+- **draw 100 lines** is 1.54x faster than draw 100 text lines (standard font)
+- **draw 100 lines** is 2.07x faster than create 10 pages with mixed content
+
+## Forms
+
+| Benchmark         | ops/sec |   Mean |    p99 |    RME | Samples |
+| :---------------- | ------: | -----: | -----: | -----: | ------: |
+| read field values |   702.6 | 1.42ms | 1.85ms | ±0.97% |     352 |
+| get form fields   |   677.3 | 1.48ms | 2.36ms | ±1.45% |     339 |
+| flatten form      |   198.8 | 5.03ms | 5.84ms | ±1.28% |     100 |
+| fill text fields  |   155.1 | 6.45ms | 7.36ms | ±1.25% |      78 |
+
+- **read field values** is 1.04x faster than get form fields
+- **read field values** is 3.53x faster than flatten form
+- **read field values** is 4.53x faster than fill text fields
+
+## Loading
+
+| Benchmark              | ops/sec |   Mean |    p99 |    RME | Samples |
+| :--------------------- | ------: | -----: | -----: | -----: | ------: |
+| load small PDF (888B)  |   38.7K |   26us |   37us | ±1.10% |  19,336 |
+| load medium PDF (19KB) |   23.8K |   42us |   54us | ±0.84% |  11,904 |
+| load form PDF (116KB)  |    1.6K |  639us | 1.07ms | ±0.96% |     782 |
+| load heavy PDF (9.9MB) |   909.7 | 1.10ms | 1.46ms | ±0.81% |     455 |
+
+- **load small PDF (888B)** is 1.62x faster than load medium PDF (19KB)
+- **load small PDF (888B)** is 24.73x faster than load form PDF (116KB)
+- **load small PDF (888B)** is 42.51x faster than load heavy PDF (9.9MB)
+
+## Saving
+
+| Benchmark                          | ops/sec |   Mean |    p99 |    RME | Samples |
+| :--------------------------------- | ------: | -----: | -----: | -----: | ------: |
+| save unmodified (19KB)             |   22.2K |   45us |   63us | ±0.78% |  11,103 |
+| incremental save (19KB)            |    6.9K |  144us |  373us | ±1.14% |   3,461 |
+| save with modifications (19KB)     |    2.4K |  422us |  799us | ±0.97% |   1,185 |
+| save heavy PDF (9.9MB)             |   850.9 | 1.18ms | 1.58ms | ±0.91% |     426 |
+| incremental save heavy PDF (9.9MB) |   494.7 | 2.02ms | 2.37ms | ±0.74% |     248 |
+
+- **save unmodified (19KB)** is 3.21x faster than incremental save (19KB)
+- **save unmodified (19KB)** is 9.37x faster than save with modifications (19KB)
+- **save unmodified (19KB)** is 26.10x faster than save heavy PDF (9.9MB)
+- **save unmodified (19KB)** is 44.89x faster than incremental save heavy PDF (9.9MB)
+
+## Splitting
+
+### Extract single page
+
+| Benchmark                                | ops/sec |    Mean |     p99 |    RME | Samples |
+| :--------------------------------------- | ------: | ------: | ------: | -----: | ------: |
+| extractPages (1 page from small PDF)     |    2.2K |   452us |   931us | ±1.63% |   1,106 |
+| extractPages (1 page from 100-page PDF)  |   536.3 |  1.86ms |  3.04ms | ±1.54% |     269 |
+| extractPages (1 page from 2000-page PDF) |    24.7 | 40.43ms | 42.46ms | ±2.46% |      13 |
+
+- **extractPages (1 page from small PDF)** is 4.12x faster than extractPages (1 page from 100-page PDF)
+- **extractPages (1 page from small PDF)** is 89.36x faster than extractPages (1 page from 2000-page PDF)
+
+### Split into single-page PDFs
+
+| Benchmark                   | ops/sec |     Mean |      p99 |    RME | Samples |
+| :-------------------------- | ------: | -------: | -------: | -----: | ------: |
+| split 100-page PDF (0.1MB)  |    32.2 |  31.02ms |  35.44ms | ±2.72% |      17 |
+| split 2000-page PDF (0.9MB) |     1.8 | 550.66ms | 550.66ms | ±0.00% |       1 |
+
+- **split 100-page PDF (0.1MB)** is 17.75x faster than split 2000-page PDF (0.9MB)
+
+### Batch page extraction
+
+| Benchmark                                              | ops/sec |    Mean |     p99 |    RME | Samples |
+| :----------------------------------------------------- | ------: | ------: | ------: | -----: | ------: |
+| extract first 10 pages from 2000-page PDF              |    23.3 | 42.90ms | 46.80ms | ±3.59% |      12 |
+| extract first 100 pages from 2000-page PDF             |    20.1 | 49.79ms | 52.72ms | ±2.98% |      11 |
+| extract every 10th page from 2000-page PDF (200 pages) |    18.6 | 53.74ms | 59.30ms | ±3.19% |      10 |
+
+- **extract first 10 pages from 2000-page PDF** is 1.16x faster than extract first 100 pages from 2000-page PDF
+- **extract first 10 pages from 2000-page PDF** is 1.25x faster than extract every 10th page from 2000-page PDF (200 pages)
+
+---
+
+_Results are machine-dependent. Use for relative comparison only._
diff --git a/scripts/bench-report.ts b/scripts/bench-report.ts
new file mode 100644
index 0000000..d16617d
--- /dev/null
+++ b/scripts/bench-report.ts
@@ -0,0 +1,262 @@
+/**
+ * Benchmark report generator.
+ *
+ * Runs vitest bench with JSON output, then transforms the results
+ * into a markdown report saved to reports/benchmarks.md.
+ *
+ * Usage:
+ *   bun run scripts/bench-report.ts
+ *   bun run scripts/bench-report.ts --json-only    # Just dump JSON, skip markdown
+ *   bun run scripts/bench-report.ts --from-json results.json  # Generate from existing JSON
+ */
+
+import { execSync } from "node:child_process";
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { cpus, platform, totalmem } from "node:os";
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Types for vitest bench JSON output
+// ─────────────────────────────────────────────────────────────────────────────
+
+interface BenchmarkResult {
+  name: string;
+  rank: number;
+  rme: number;
+  hz: number;
+  min: number;
+  max: number;
+  mean: number;
+  p75: number;
+  p99: number;
+  p995: number;
+  p999: number;
+  sampleCount: number;
+  median: number;
+}
+
+interface BenchmarkGroup {
+  fullName: string;
+  benchmarks: BenchmarkResult[];
+}
+
+interface BenchmarkFile {
+  filepath: string;
+  groups: BenchmarkGroup[];
+}
+
+interface BenchmarkOutput {
+  files: BenchmarkFile[];
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Helpers
+// ─────────────────────────────────────────────────────────────────────────────
+
+function formatHz(hz: number): string {
+  if (hz >= 1000) {
+    return `${(hz / 1000).toFixed(1)}K`;
+  }
+
+  if (hz >= 1) {
+    return hz.toFixed(1);
+  }
+
+  return hz.toFixed(3);
+}
+
+function formatTime(ms: number): string {
+  if (ms >= 1000) {
+    return `${(ms / 1000).toFixed(2)}s`;
+  }
+
+  if (ms >= 1) {
+    return `${ms.toFixed(2)}ms`;
+  }
+
+  return `${(ms * 1000).toFixed(0)}us`;
+}
+
+function formatRme(rme: number): string {
+  return `\u00b1${rme.toFixed(2)}%`;
+}
+
+function getSystemInfo(): string {
+  const cpu = cpus()[0];
+  const cpuModel = cpu?.model ?? "Unknown CPU";
+  const cpuCount = cpus().length;
+  const mem = (totalmem() / 1024 / 1024 / 1024).toFixed(0);
+  const os = platform();
+  const runtime = `Bun ${process.versions.bun ?? "unknown"}`;
+
+  return `${os} | ${cpuModel} (${cpuCount} cores) | ${mem}GB RAM | ${runtime}`;
+}
+
+/**
+ * Extract a short file label from a benchmark filepath.
+ * e.g. "/Users/.../benchmarks/loading.bench.ts" -> "Loading"
+ */
+function fileLabel(filepath: string): string {
+  const match = filepath.match(/([^/]+)\.bench\.ts$/);
+
+  if (!match) {
+    return filepath;
+  }
+
+  const name = match[1];
+
+  return name.charAt(0).toUpperCase() + name.slice(1);
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Markdown generation
+// ─────────────────────────────────────────────────────────────────────────────
+
+function generateMarkdown(data: BenchmarkOutput): string {
+  const lines: string[] = [];
+  const now = new Date();
+  const dateStr = now.toISOString().split("T")[0];
+  const timeStr = now.toISOString().split("T")[1].split(".")[0];
+
+  lines.push("# Benchmark Report");
+  lines.push("");
+  lines.push(`> Generated on ${dateStr} at ${timeStr} UTC`);
+  lines.push(`>`);
+  lines.push(`> System: ${getSystemInfo()}`);
+  lines.push("");
+  lines.push("---");
+  lines.push("");
+
+  // Table of contents
+  lines.push("## Contents");
+  lines.push("");
+
+  for (const file of data.files) {
+    const label = fileLabel(file.filepath);
+    const anchor = label.toLowerCase().replace(/\s+/g, "-");
+    lines.push(`- [${label}](#${anchor})`);
+  }
+
+  lines.push("");
+
+  // Each file becomes a section
+  for (const file of data.files) {
+    const label = fileLabel(file.filepath);
+    lines.push(`## ${label}`);
+    lines.push("");
+
+    for (const group of file.groups) {
+      // If the group name differs from the file-level name, add a subheading
+      const groupName = group.fullName.replace(/^benchmarks\/[^>]+> /, "").trim();
+      const isTopLevel = group.benchmarks.length > 0 && !groupName.includes(" > ");
+
+      // Check if this group has a describe() wrapper (indicated by " > " in fullName)
+      const describeName = group.fullName.includes(" > ")
+        ? group.fullName.split(" > ").slice(1).join(" > ")
+        : null;
+
+      if (describeName) {
+        lines.push(`### ${describeName}`);
+        lines.push("");
+      }
+
+      // Build the results table
+      lines.push("| Benchmark | ops/sec | Mean | p99 | RME | Samples |");
+      lines.push("|:---|---:|---:|---:|---:|---:|");
+
+      // Sort by rank
+      const sorted = [...group.benchmarks].sort((a, b) => a.rank - b.rank);
+
+      for (const bench of sorted) {
+        const name = bench.name;
+        const hz = formatHz(bench.hz);
+        const mean = formatTime(bench.mean);
+        const p99 = formatTime(bench.p99);
+        const rme = formatRme(bench.rme);
+        const samples = bench.sampleCount.toLocaleString();
+
+        lines.push(`| ${name} | ${hz} | ${mean} | ${p99} | ${rme} | ${samples} |`);
+      }
+
+      lines.push("");
+
+      // Add comparison summary for groups with multiple benchmarks
+      if (sorted.length >= 2) {
+        const fastest = sorted[0];
+        const rest = sorted.slice(1);
+
+        for (const slower of rest) {
+          const ratio = (fastest.hz / slower.hz).toFixed(2);
+          lines.push(`- **${fastest.name}** is ${ratio}x faster than ${slower.name}`);
+        }
+
+        lines.push("");
+      }
+    }
+  }
+
+  // Footer
+  lines.push("---");
+  lines.push("");
+  lines.push("*Results are machine-dependent. Use for relative comparison only.*");
+  lines.push("");
+
+  return lines.join("\n");
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Main
+// ─────────────────────────────────────────────────────────────────────────────
+
+const args = process.argv.slice(2);
+const jsonOnly = args.includes("--json-only");
+const fromJsonIdx = args.indexOf("--from-json");
+
+const jsonPath = "reports/bench-results.json";
+const mdPath = "reports/benchmarks.md";
+
+mkdirSync("reports", { recursive: true });
+
+let data: BenchmarkOutput;
+
+if (fromJsonIdx !== -1 && args[fromJsonIdx + 1]) {
+  // Generate markdown from an existing JSON file
+  const inputPath = args[fromJsonIdx + 1];
+
+  if (!existsSync(inputPath)) {
+    console.error(`File not found: ${inputPath}`);
+    process.exit(1);
+  }
+
+  data = JSON.parse(readFileSync(inputPath, "utf-8")) as BenchmarkOutput;
+  console.log(`Loaded benchmark results from ${inputPath}`);
+} else {
+  // Run benchmarks and capture JSON
+  console.log("Running benchmarks...\n");
+
+  try {
+    execSync(`bun run bench -- --outputJson ${jsonPath}`, {
+      stdio: "inherit",
+      timeout: 600_000, // 10 minute timeout
+    });
+  } catch (error) {
+    console.error("Benchmark run failed");
+    process.exit(1);
+  }
+
+  if (!existsSync(jsonPath)) {
+    console.error(`Expected JSON output at ${jsonPath} but file not found`);
+    process.exit(1);
+  }
+
+  data = JSON.parse(readFileSync(jsonPath, "utf-8")) as BenchmarkOutput;
+  console.log(`\nBenchmark JSON saved to ${jsonPath}`);
+}
+
+if (jsonOnly) {
+  process.exit(0);
+}
+
+// Generate and write markdown report
+const md = generateMarkdown(data);
+writeFileSync(mdPath, md);
+console.log(`Benchmark report saved to ${mdPath}`);