Kernel is #1 on ComputeSDK's browser throughput benchmark

Kernel ranks #1 on ComputeSDK's browser throughput benchmark as of July 15, 2026, with a composite score of 77.5. ComputeSDK is an independent, daily benchmark of every major cloud browser infrastructure provider, run by a third party and published in an open-source GitHub repository.

Key findings: Kernel's benchmark results

• Kernel ranks #1 with a composite score of 77.5, ahead of Browser Use (70.1), Browserbase (61.6), Steel (52.1), Hyperbrowser (52.0), and Notte (48.6).
• Kernel executes 5.28 actions per second at the median, roughly 2.3x Browserbase's 2.29 actions per second.
• Median task duration on Kernel is 2.13 seconds, versus 3.41 seconds on Browser Use and 4.90 seconds on Browserbase.
• Kernel posts the lowest individual action times in the benchmark: 253ms screenshots, 549ms navigation, 204ms clicks.
• Kernel's p95 task duration of 2.89s sits close to its median of 1.89s, indicating consistent performance for AI agent workloads.

About the ComputeSDK browser throughput benchmark

The ComputeSDK browser throughput benchmark is an independent, open-source performance test that runs daily against cloud browser infrastructure providers. It's published at computesdk.com/benchmarks/browsers/browser-throughput, with methodology and raw results in the computesdk/benchmarks GitHub repository.

Most vendor-published benchmarks are worth what you paid for them. The ComputeSDK benchmark is worth attention for four reasons:

• It's independent. Kernel did not design, run, or fund it.
• It runs daily via GitHub Actions, so results reflect current performance rather than a one-time snapshot.
• Methodology and raw JSON data are public and reproducible.
• The composite score is multiplied by success rate. A provider that's fast but flaky is penalized proportionally, so 1-in-10 failures can't hide behind a fast average.

How the benchmark works

Each provider runs 10 sessions of 50 sequential actions inside a single running headless browser. The action loop is: navigate, waitForSelector, screenshot, textContent, click, waitForSelector, screenshot, textContent, goBack, waitForSelector. That 10-step sequence repeats 5 times per session, run against Wikipedia. All sessions use stealth mode at a 1920×1080 viewport to approximate real AI agent conditions.

The composite score (0-100, higher is better) weights four axes:

• Actions per second median: 40%
• Task duration median: 25%
• Task duration p95: 20%
• Screenshot median: 15%

The final score is multiplied by success rate.

Full results: cloud browser provider rankings

How Kernel's throughput and latency numbers translate to agent workloads

Actions per second is the throughput ceiling for AI agents. If an agent takes 20 actions to complete a task at 5 actions per second, that's 4 seconds of infrastructure overhead per task. At 2 actions per second, it's 10 seconds. That gap compounds across every task an agent runs.

Tail latency (p95) matters more than the average for long-running sessions. A fast average with a slow tail means most sessions feel snappy and every tenth one times out. Kernel's p95 of 2.89s sits close to its median of 1.89s. Steel is at 5.56s median and 13.67s p95, where the tail is more than twice the median.

Per-action times compound across a task. A vision-based agent takes a screenshot after every action. The difference between 250ms and 400ms per screenshot doesn't sound like much until you multiply it by a 50-step task.

What this benchmark doesn't measure

Worth being upfront about scope:

• No hard anti-bot targets. The benchmark runs against Wikipedia. Workloads involving Cloudflare Turnstile, DataDome, PerimeterX, or Akamai Bot Manager are a different test.
• No long-running sessions. 50 actions is not a 30-minute research task.
• No concurrency stress. This measures throughput inside a session, not how a provider handles 1,000 sessions launching at once.

If any of those are the shape of your workload, the ComputeSDK repo is a solid starting point for adding your own tests. Or get in touch and we can talk about your specific case on Kernel.

Frequently asked questions

Which cloud browser infrastructure provider is fastest for AI agents?

As of July 13, 2026, Kernel ranks #1 on ComputeSDK's independent browser throughput benchmark with a composite score of 77.5. Kernel leads on actions per second, median task duration, p95 tail latency, and every individual action type measured (screenshot, navigate, click).

How does Kernel compare to Browserbase?

On the ComputeSDK browser throughput benchmark (July 13, 2026), Kernel scores 77.5 versus Browserbase's 61.6. Kernel executes 5.28 actions per second at the median versus Browserbase's 2.29, roughly 2.3x the throughput. Median task duration is 2.13 seconds on Kernel versus 4.90 seconds on Browserbase.

How does Kernel compare to Browser Use?

On the ComputeSDK browser throughput benchmark (July 13, 2026), Kernel scores 77.5 versus Browser Use's 70.1. Kernel executes 5.28 actions per second at the median versus Browser Use's 3.77, roughly 1.4x the throughput. Median task duration is 2.13 seconds on Kernel versus 3.41 seconds on Browser Use.

What is the ComputeSDK browser throughput benchmark?

The ComputeSDK browser throughput benchmark is an independent, open-source performance test that runs daily against major cloud browser infrastructure providers. It measures actions per second, task duration, tail latency, and per-action timings across 10 sessions of 50 sequential actions per provider. Methodology and raw data are public on GitHub at computesdk/benchmarks.

Why does actions per second matter for browser agents?

Actions per second is the throughput ceiling for how fast an AI agent can iterate through a task. Higher actions per second means less infrastructure overhead per task, which compounds across every action an agent takes. At 5 actions per second, a 20-step task takes 4 seconds of infrastructure time. At 2 actions per second, the same task takes 10 seconds.

Is Kernel independent of ComputeSDK?

Yes. Kernel did not design, run, or fund the ComputeSDK benchmark. ComputeSDK runs the benchmark independently via GitHub Actions, and all methodology and raw data are publicly available in the computesdk/benchmarks repository.

Related resources

• Kernel: browser infrastructure for AI agents
• ComputeSDK browser throughput benchmark, full leaderboard
• ComputeSDK benchmarks source code and raw data on GitHub

Last updated: July 14, 2026. Benchmark results reflect the ComputeSDK daily run on July 13, 2026. For current standings, see the ComputeSDK benchmark page.