diff --git a/.github/scripts/compare-benchmarks.py b/.github/scripts/compare-benchmarks.py
new file mode 100755
index 0000000..e513f1a
--- /dev/null
+++ b/.github/scripts/compare-benchmarks.py
@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+"""Compare two JMH JSON result files, print a Markdown table.
+
+Used by benchmark-lto.yml to summarize LTO's throughput impact per platform:
+compare-benchmarks.py <baseline.json> <variant.json>
+"""
+import json
+import sys
+
+
+def load(path):
+    with open(path) as f:
+        results = json.load(f)
+    return {
+        (r["benchmark"], tuple(sorted(r.get("params", {}).items()))): r
+        for r in results
+    }
+
+
+def main():
+    if len(sys.argv) != 3:
+        print("usage: compare-benchmarks.py <baseline.json> <variant.json>", file=sys.stderr)
+        return 1
+
+    baseline = load(sys.argv[1])
+    variant = load(sys.argv[2])
+
+    print("| Benchmark | Params | Baseline | Variant | Delta |")
+    print("|---|---|---:|---:|---:|")
+    for key in sorted(baseline):
+        if key not in variant:
+            continue
+        b, v = baseline[key]["primaryMetric"], variant[key]["primaryMetric"]
+        # scoreError is "NaN" (a JSON string, not a number) when JMH can't
+        # compute a confidence interval - e.g. a single measurement iteration.
+        b_score, b_err, unit = b["score"], float(b["scoreError"]), b["scoreUnit"]
+        v_score, v_err = v["score"], float(v["scoreError"])
+        delta = (v_score - b_score) / b_score * 100 if b_score else 0.0
+        name = key[0].rsplit(".", 1)[-1]
+        params = ", ".join(f"{k}={pv}" for k, pv in key[1]) or "-"
+        print(
+            f"| {name} | {params} | {b_score:.3f} ± {b_err:.3f} {unit} "
+            f"| {v_score:.3f} ± {v_err:.3f} {unit} | {delta:+.1f}% |"
+        )
+
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/.github/workflows/benchmark-lto.yml b/.github/workflows/benchmark-lto.yml
new file mode 100644
index 0000000..3541a98
--- /dev/null
+++ b/.github/workflows/benchmark-lto.yml
@@ -0,0 +1,86 @@
+name: Benchmark LTO impact
+
+# Manual, exploratory: compares JMH throughput between main and an LTO
+# variant branch (see scripts/build-zstd.sh - currently -flto is Linux-only,
+# zig's Mach-O backend doesn't support it) on every runner OS the repo
+# supports, so the real-vs-no-op split shows up as CI evidence per platform
+# instead of a manual claim. Not part of ci.yml: this is throughput
+# exploration, not a correctness gate, and JMH runs are too slow to run on
+# every push/PR.
+
+on:
+  workflow_dispatch:
+    inputs:
+      lto_ref:
+        description: Branch/ref with the LTO change to compare against main.
+        required: false
+        default: experiment/lto-linux
+
+jobs:
+  benchmark:
+    name: ${{ matrix.classifier }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { os: ubuntu-latest,    classifier: linux-x86_64 }
+          - { os: ubuntu-24.04-arm, classifier: linux-aarch64 }
+          - { os: macos-14,         classifier: osx-aarch64 }
+          - { os: windows-latest,   classifier: windows-x86_64 }
+    runs-on: ${{ matrix.os }}
+    steps:
+      - name: Checkout baseline (main, with zstd submodule)
+        uses: actions/checkout@v7
+        with:
+          submodules: recursive
+          path: baseline
+
+      - name: Checkout LTO variant (${{ inputs.lto_ref }}, with zstd submodule)
+        uses: actions/checkout@v7
+        with:
+          submodules: recursive
+          ref: ${{ inputs.lto_ref }}
+          path: lto
+
+      - name: Set up JDK 25
+        uses: actions/setup-java@v5
+        with:
+          distribution: temurin
+          java-version: '25'
+          cache: maven
+
+      - name: Set up Zig
+        uses: mlugg/setup-zig@d1434d08867e3ee9daa34448df10607b98908d29 # v2
+        with:
+          version: 0.16.0
+
+      # `ompressBenchmark` matches both CompressBenchmark and
+      # DecompressBenchmark method FQNs (JMH's filter is a substring/regex
+      # search), but not GoldenCorpusBenchmark - keeps this to the quick
+      # synthetic-payload suite, not the slower real-corpus one.
+      - name: Build + run baseline benchmark
+        shell: bash
+        working-directory: baseline
+        run: |
+          ./mvnw -B -ntp -q -pl benchmark -am package -DskipTests
+          java -jar benchmark/target/benchmarks.jar ompressBenchmark \
+            -f 1 -wi 2 -i 5 -p size=65536 -rf json -rff ../baseline-results.json
+
+      - name: Build + run LTO-variant benchmark
+        shell: bash
+        working-directory: lto
+        run: |
+          ./mvnw -B -ntp -q -pl benchmark -am package -DskipTests
+          java -jar benchmark/target/benchmarks.jar ompressBenchmark \
+            -f 1 -wi 2 -i 5 -p size=65536 -rf json -rff ../lto-results.json
+
+      - name: Summarize
+        shell: bash
+        run: |
+          {
+            echo "### ${{ matrix.classifier }}"
+            echo
+            python3 baseline/.github/scripts/compare-benchmarks.py \
+              baseline-results.json lto-results.json
+            echo
+          } >> "$GITHUB_STEP_SUMMARY"