diff --git a/maintainers/eval-benchmark.py b/maintainers/eval-benchmark.py
new file mode 100755
index 000000000000..b7a796e3c960
--- /dev/null
+++ b/maintainers/eval-benchmark.py
@@ -0,0 +1,134 @@
+#!/usr/bin/env nix
+#!nix shell --inputs-from .. nixpkgs#python3 --command python3
+
+"""
+Benchmark Nix evaluation performance across a set of Nix releases.
+
+For each tag (e.g. `v3.23.1` for Determinate Nix, `2.35.2` for upstream Nix,
+or a full Git revision of DeterminateSystems/nix-src), each test and each run
+number, this runs the test and appends a row to the CSV file. Runs that are
+already recorded in the CSV file are skipped, and Nix versions are only built
+when needed.
+"""
+
+import argparse
+import csv
+import os
+import re
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+NIXPKGS = "github:NixOS/nixpkgs/531670d871c0e29724a02f3cbcac170adc65b58c"
+
+TESTS = {
+    "search": ["search", NIXPKGS, "fizzbuzz", "--no-eval-cache"],
+    "firefox": ["eval", "--json", f"{NIXPKGS}#firefox", "--read-only"],
+    "plasma": [
+        "eval",
+        "--json",
+        "-I",
+        f"nixpkgs=flake:{NIXPKGS}",
+        "--file",
+        "<nixpkgs/nixos/release-combined.nix>",
+        "nixos.tests.plasma6.x86_64-linux",
+        "--read-only",
+    ],
+}
+
+EXTRA_ARGS = ["--option", "eval-cores", "0"]
+
+HEADER = ["nix-tag", "test-name", "test-run", "elapsed-time", "cpu-time", "kernel-time", "max-rss-kib"]
+
+
+def flake_ref(tag: str) -> str:
+    # Determinate Nix tags and revisions come from nix-src, upstream tags from NixOS/nix.
+    if tag.startswith("v") or re.fullmatch(r"[0-9a-f]{40}", tag):
+        return f"github:DeterminateSystems/nix-src/{tag}"
+    return f"github:NixOS/nix/{tag}"
+
+
+def build_nix(tag: str) -> Path | None:
+    ref = flake_ref(tag)
+    print(f"building {ref}...", file=sys.stderr)
+    res = subprocess.run(["nix", "build", "--no-link", "--print-out-paths", ref], stdout=subprocess.PIPE, text=True)
+    if res.returncode != 0:
+        print(f"failed to build {ref}, skipping", file=sys.stderr)
+        return None
+    # The package has multiple outputs (e.g. `man`), so find the one with the binary.
+    for out in res.stdout.split():
+        nix_bin = Path(out) / "bin" / "nix"
+        if os.access(nix_bin, os.X_OK):
+            return nix_bin
+    print(f"{ref} does not provide bin/nix, skipping", file=sys.stderr)
+    return None
+
+
+def run_test(nix_bin: Path, test: str) -> list[str] | None:
+    """Run a test, returning the measurements, or None if the test failed."""
+    start = time.monotonic()
+    proc = subprocess.Popen([nix_bin, *TESTS[test], *EXTRA_ARGS], stdout=subprocess.DEVNULL)
+    _, status, rusage = os.wait4(proc.pid, 0)
+    elapsed = time.monotonic() - start
+    proc.returncode = os.waitstatus_to_exitcode(status)
+    if proc.returncode != 0:
+        return None
+    return [
+        f"{elapsed:.3f}",
+        f"{rusage.ru_utime:.3f}",
+        f"{rusage.ru_stime:.3f}",
+        str(rusage.ru_maxrss),  # KiB on Linux
+    ]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("tags", metavar="TAG", nargs="+", help="Nix release tag or nix-src revision")
+    parser.add_argument("-o", "--output", type=Path, default=Path("eval-benchmark.csv"), help="CSV file to append to")
+    parser.add_argument("-n", "--runs", type=int, default=5, help="number of runs per test/tag pair")
+    parser.add_argument("-t", "--test", dest="tests", action="append", choices=TESTS.keys(), help="test to run (default: all)")
+    args = parser.parse_args()
+
+    tests = args.tests or list(TESTS.keys())
+
+    done = set()
+    if args.output.exists():
+        with args.output.open(newline="") as f:
+            for row in csv.DictReader(f):
+                done.add((row["nix-tag"], row["test-name"], row["test-run"]))
+    else:
+        with args.output.open("w", newline="") as f:
+            csv.writer(f).writerow(HEADER)
+
+    for tag in args.tags:
+        nix_bin = None
+        build_failed = False
+
+        for test in tests:
+            if build_failed:
+                break
+
+            for run in range(1, args.runs + 1):
+                if (tag, test, str(run)) in done:
+                    continue
+
+                if nix_bin is None:
+                    nix_bin = build_nix(tag)
+                    if nix_bin is None:
+                        build_failed = True
+                        break
+
+                measurements = run_test(nix_bin, test)
+                if measurements is None:
+                    print(f"{tag} {test} {run}: failed", file=sys.stderr)
+                    break
+
+                # Append each row immediately so that an interrupted run loses at most one measurement.
+                with args.output.open("a", newline="") as f:
+                    csv.writer(f).writerow([tag, test, run, *measurements])
+                print(f"{tag} {test} {run}: {measurements[0]} s", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/maintainers/plot-eval-benchmark.py b/maintainers/plot-eval-benchmark.py
new file mode 100755
index 000000000000..5e4e4da4e54b
--- /dev/null
+++ b/maintainers/plot-eval-benchmark.py
@@ -0,0 +1,124 @@
+#!/usr/bin/env nix
+#! nix shell --impure --expr ``
+#! nix with (builtins.getFlake (toString ./..)).inputs.nixpkgs.legacyPackages.${builtins.currentSystem};
+#! nix python3.withPackages (ps: [ ps.matplotlib ])
+#! nix ``
+#! nix --command python3
+
+"""Plot the minimum elapsed time and max RSS per Nix release for each test in an eval-benchmark.py CSV file."""
+
+import argparse
+import csv
+import re
+from collections import defaultdict
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument("csv", nargs="?", default="eval-benchmark.csv", metavar="FILE.csv")
+parser.add_argument(
+    "-o", "--output", help="output file (default: eval-benchmark.png, or .svg with --svg)"
+)
+parser.add_argument(
+    "--svg",
+    action="store_true",
+    help="write a transparent SVG styled for a dark background",
+)
+parser.add_argument("--log", action="store_true", help="use a logarithmic y-axis")
+args = parser.parse_args()
+
+if args.output is None:
+    args.output = "eval-benchmark.svg" if args.svg else "eval-benchmark.png"
+
+if args.svg:
+    fg = "#dddddd"
+    plt.rcParams.update({
+        "svg.fonttype": "none",
+        "text.color": fg,
+        "axes.labelcolor": fg,
+        "axes.edgecolor": fg,
+        "axes.titlecolor": fg,
+        "xtick.color": fg,
+        "ytick.color": fg,
+        "grid.color": fg,
+        "legend.edgecolor": fg,
+        "legend.facecolor": "none",
+        "legend.labelcolor": fg,
+        "figure.facecolor": "none",
+        "axes.facecolor": "none",
+        "savefig.facecolor": "none",
+        "savefig.edgecolor": "none",
+    })
+
+
+def version_key(tag):
+    """Sort key for tags such that v3.1.0 < v3.2.0 < v3.10.0."""
+    return [int(p) if p.isdigit() else p for p in re.split(r"(\d+)", tag.removeprefix("v"))]
+
+
+def is_revision(tag):
+    return re.fullmatch(r"[0-9a-f]{40}", tag) is not None
+
+
+def is_determinate(tag):
+    # Revisions are of DeterminateSystems/nix-src.
+    return tag.startswith("v") or is_revision(tag)
+
+
+# (CSV column, y-axis label, scale factor), plotted as columns from left to right.
+METRICS = [
+    ("elapsed-time", "Elapsed time (s)", 1),
+    ("max-rss-kib", "Max RSS (MiB)", 1 / 1024),
+]
+
+# data[metric][test][tag] = list of measurements
+data = {metric: defaultdict(lambda: defaultdict(list)) for metric, _, _ in METRICS}
+with open(args.csv, newline="") as f:
+    for row in csv.DictReader(f):
+        for metric, _, scale in METRICS:
+            data[metric][row["test-name"]][row["nix-tag"]].append(float(row[metric]) * scale)
+
+times = data["elapsed-time"]
+tests = list(times)
+# Release tags sorted by version, followed by revisions in the order in which they appear in the CSV file.
+all_tags = list(dict.fromkeys(tag for per_tag in times.values() for tag in per_tag))
+tags = sorted([t for t in all_tags if not is_revision(t)], key=version_key) + [t for t in all_tags if is_revision(t)]
+x = {tag: i for i, tag in enumerate(tags)}
+
+fig, axes = plt.subplots(
+    len(tests),
+    len(METRICS),
+    figsize=(max(10, 0.4 * len(tags)) * len(METRICS), 4 * len(tests)),
+    sharex=True,
+    squeeze=False,
+)
+
+for row_axes, test in zip(axes, tests):
+    for ax, (metric, ylabel, _) in zip(row_axes, METRICS):
+        values = data[metric][test]
+        for name, pred in [("Nix", lambda t: not is_determinate(t)), ("Determinate Nix", is_determinate)]:
+            series = [tag for tag in tags if tag in values and pred(tag)]
+            if series:
+                ax.plot([x[t] for t in series], [min(values[t]) for t in series], marker=".", label=name)
+        ax.set_title(test)
+        ax.set_ylabel(ylabel)
+        if args.log:
+            ax.set_yscale("log")
+        else:
+            ax.set_ylim(bottom=0)
+        ax.grid(True, axis="y", linestyle=":", alpha=0.5)
+        ax.legend(title="minimum values")
+
+for ax in axes[-1]:
+    ax.set_xticks(range(len(tags)), [t[:10] if is_revision(t) else t for t in tags], rotation=45, ha="right")
+    ax.set_xlabel("Nix release")
+fig.suptitle("Nix evaluation performance")
+fig.tight_layout()
+if args.svg:
+    fig.savefig(args.output, format="svg", transparent=True)
+else:
+    fig.savefig(args.output, dpi=150)
+print(f"wrote {args.output}")