baton

Benchmarks

Measured 27 Oct 2026 from a client in us-east-1, on the same machines that serve this site. Every figure here is read from the raw samples, which are public.

475 ms

a whole session on Browser Arena

6th of 13 by the median, 4th by p95 (564 ms)

222 ms

new browser, request → answering CDP

p95 273 ms

223 ms

paused while it moves machines

a live client’s longest wait: 595 ms

201 ms

paused to fork a running browser

3 copies ready in 2.6 s

Browser Arena

The harness behind browserarena.ai with Baton added as a provider and nothing else changed: create, connect over CDP, load example.com, release. 10 warm-ups, then 100 sessions one at a time, from a t3.small in us-east-1, on a free account.

Browser Arena

Milliseconds per session, lower is better

  1. N

    Notte

    340
    p50291p99430
  2. Browser Use

    490
    p50424p99639
  3. L

    Lightpanda

    559
    p50485p99674
  4. Baton

    564
    p50475p99630
  5. Browserbase

    663
    p50590p99730
  6. Kernel (headful)

    676
    p50344p99767
  7. Kernel

    712
    p50378p99741
  8. Tilion

    922
    p50156p9913,292
  9. Hyperbrowser

    975
    p50890p991,280
  10. Cloudflare

    1,546
    p50962p991,919
  11. Steel

    1,691
    p501,293p992,577
  12. Anchor

    3,164
    p502,552p993,937
  13. AWS AgentCore

    3,627
    p502,613p994,411

Every other row is Browser Arena’s own latest run, from the raw results it publishes; we did not re-run them. Baton’s row is our run of the same harness from t3.small in us-east-1, on a free account, 100 of 100 sessions completed.

One day of work on it, 2026-10-24

the sum of the stage medians, as Browser Arena reports it

createconnectload pagereleasetotalp95
Start of the day244 ms259 ms489 ms206 ms1,198 ms1,331 ms
Release in the background232 ms262 ms493 ms51 ms1,038 ms1,148 ms
Warm pool70 ms122 ms323 ms53 ms568 ms688 ms
One-hop API58 ms99 ms270 ms43 ms470 ms564 ms

Where the rest goes, and why we stopped here

  • Sessions start from a browser that is already running and warmed up, a pool of one per machine; claiming one is a database update. That took create from 244 ms to 58 ms and connect from 259 ms to 99 ms.
  • Loading the page is most of what is left, and it is Chrome, not the network: our gateway fetches example.com in about 25 ms. A browser in a VM on nested virtualization pays for every first touch of a memory page (12–17 µs measured) and every wake-up across CPUs. The fastest provider here loads the page in about 46 ms.
  • About 40 ms of connect is CloudFront, which gives us TLS without a domain of our own.
  • Getting near 150 ms would take bare-metal hosts (EC2 bare metal starts around $4 an hour per machine, always on) and a TLS endpoint of our own. That is spending, not engineering, so we stopped.
  • The pool’s costs: one idle browser per machine, and a warmed browser has touched more memory, so pausing one stores about 740 MiB instead of 610 (pause 2.2 s instead of 1.9).

Steel’s browserbench

Steel’s open-source benchmark with Baton added: create, connect, load google.com, release; 300 runs (300 succeeded) from a t3.small in us-east-1 on October 24, before the warm pool and the one-hop API. The other rows are the numbers Steel published; we did not re-run them.

Milliseconds per session

total is the median; stages are averages, as in Steel’s table

total p50p95createconnectload pagereleaseall but page
Steel867 ms1,090 ms182 ms175 ms490 ms48 ms404 ms
Kernel1,485 ms1,752 ms261 ms429 ms534 ms295 ms985 ms
Browserbase1,669 ms1,873 ms188 ms517 ms792 ms179 ms885 ms
Baton2,084 ms2,191 ms245 ms271 ms1,358 ms219 ms735 ms
Hyperbrowser3,666 ms5,338 ms1,732 ms348 ms378 ms1,200 ms3,279 ms
AnchorBrowser7,919 ms11,561 ms3,797 ms184 ms1,260 ms2,761 ms6,742 ms
  • Without the page load, a Baton session cost 735 ms: second-fastest of the six.
  • Loading google.com is where Baton loses: 1,358 ms against 378–1,260 ms. The network is not the cause (first byte after about 75 ms). A browser restored from a snapshot runs cold: its first load of a heavy page takes about 1.3 s, and about 350 ms by the third load. We think nested virtualization makes every first touch of a memory page slow; we have not run on bare metal.
  • Different days, instances and run lengths (Steel: 5,000 runs; ours: 300); read it as the same method, not the same hour. Raw runs: browserbench-baton.jsonl.

Moving, copying, sleeping

What the features cost the browser that has them: how long it is paused, and how long a client waits. Live migration under a running client, 0 disconnects.

Live migration

Post-copy migration between two nodes while a CDP client sends requests back to back
p50p95p99maxn
browser paused223 ms242 ms242 ms242 ms20
unreachable225 ms244 ms244 ms244 ms20
client’s longest wait595 ms669 ms669 ms669 ms20
client round trip, normally2 ms2 ms3 ms3 ms199
memory still arriving after802 ms880 ms880 ms880 ms20

Fork a running browser

A browser with a page open is copied: the copies share its cookies, storage and open pages; each gets its own address and fresh random state.
p50p95p99maxn
parent paused, 1 copy201 ms205 ms205 ms205 ms5
1 copy ready1,448 ms1,531 ms1,531 ms1,531 ms5
parent paused, 3 copies200 ms201 ms201 ms201 ms5
3 copies ready2,607 ms2,692 ms2,692 ms2,692 ms5

Pause to storage, resume anywhere

Pause a Chrome with a page open to S3 and resume it (the node may differ each time), 10 round trips; the guest clock is set on resume
p50p95p99maxn
pause: request → stored2,120 ms2,265 ms2,265 ms2,265 ms10
pause: browser stopped409 ms416 ms416 ms416 ms10
pause: upload to S31,704 ms1,900 ms1,900 ms1,900 ms10
resume: request → answering1,217 ms1,297 ms1,297 ms1,297 ms10
resume: download from S31,016 ms1,030 ms1,030 ms1,030 ms10

Start from a saved profile

Save a logged-in browser as a profile, then start sessions from it through the public API (create request to ready; the first fetches the profile to a node) Saving took 2,013 ms; the browser was paused 202 ms; 595 MiB stored.
p50p95p99maxn
first session (fetches it)1,602 ms1,602 ms1,602 ms1,602 ms1
next sessions747 ms959 ms959 ms959 ms7

Starting a browser

From the orchestrator’s side, and from a customer’s: through the internet, Vercel and CloudFront.

Fresh start

Orchestrator create: request -> Chrome answering CDP, gateway route set (server-side time)
p50p95p99maxn
create222 ms273 ms294 ms294 ms100

Public path, as a customer

A customer in us-east-1: POST /api/v1/sessions (Vercel) -> connectOverCDP (CloudFront) -> load example.com
p50p95p99maxn
create (API)64 ms290 ms1,181 ms1,181 ms40
connect (CDP)93 ms117 ms298 ms298 ms40
first page265 ms339 ms527 ms527 ms40
total433 ms630 ms2,005 ms2,005 ms40

What a browser costs

Memory a browser really costs on its node: private is its own, PSS adds its share of pages shared with others (3 fresh browsers on one node; then one of them forked into 3 copies on that node). Each browser has 1,024 MiB of guest memory.

191 MiB

a fresh browser’s own memory

312 MiB with its share of the snapshot

105 MiB

a fork copy’s own memory

212 MiB with its share

727 MiB

fork image, shared by every copy on a machine

Noisy neighbour

A victim browser times a fixed piece of JavaScript work inside its page (performance.now(), so the client's network plays no part) every 200 ms, while a hog browser on the same 2-CPU node spins both its vCPUs. Measured with scripts/bench/noisy.mjs; the timing is in-page. Measured 2026-10-24.
setupp50p95p99n
hog idle4.7 ms4.8 ms5 ms44
hog busy, both VMs in one shared cgroup (no per-VM cgroups)4.7 ms8.3 ms17.1 ms44
hog busy, own cgroups, equal weight, no cap4.7 ms14.7 ms15.1 ms44
hog busy, capped at 1 CPU, 20 ms period (the default)4.7 ms5.7 ms11.3 ms67

How, and what to keep in mind

  • Machines: m8i.large (2 vCPU, 8 GiB) with nested virtualization; gateway t3.small. VMM: Firecracker v1.17.0 under its jailer; cgroup v2 per VM (guest + 256 MiB; capped at 1 CPU once running); 2 pre-started jailed VMMs kept ready per node. Guest: Chrome for Testing 156, 2 vCPU, 1 GiB per browser.
  • Nested virtualization (KVM inside an EC2 VM) is slower than the bare-metal hosts sandbox providers usually run on; the numbers here are likely lower there (not measured).
  • The client: bench client, t3.small, same VPC. Internal tests talk to the orchestrator and gateway on private addresses; the public test goes through the internet, Vercel and CloudFront, as a customer would.
  • Fresh start is the orchestrator’s own time from request to Chrome answering, including placement, the node starting the VM in its jail, and the network route.
  • Percentiles are over the samples listed; the raw samples are in the JSON.