Last updated: July 2026. Every figure on this page links to the published benchmark it comes from. Methodology details are stated as each publisher describes them; check the linked sources for full results.
Browser infrastructure for AI agents is measurable. The question that decides most evaluations, whether an agent session survives bot detection and completes its work on real commercial websites, has been tested three separate times by three separate publishers: Halluminate's BrowserBench, Browser Use's stealth benchmark, and AIMultiple's remote-browser comparison. Anchor Browser appears in all three. This page collects the results.
The headline figures:
-
Lowest stealth failure rate on BrowserBench: 1.7%. Anchor's stealth mode posted the lowest failure rate of all eight provider configurations tested, ahead of Hyperbrowser (2.7%), Browserbase (3.4%), and Steel (6.8% with stealth enabled). (Halluminate, live leaderboard)
-
77% on Browser Use's stealth benchmark. The highest pass rate of any browser infrastructure provider in the test apart from Browser Use's own cloud, and a 30-point margin over Steel (47%), Browserbase (42%), and Hyperbrowser (40%). (Browser Use)
-
82% composite on AIMultiple's remote-browser benchmark. The highest composite and features scores among the browser infrastructure providers agent builders typically shortlist together (Steel 72%, Browserbase 65%, Hyperbrowser 62%). (AIMultiple)
Browser
Bench: the lowest stealth failure rate of any configuration tested
BrowserBench is a public benchmark for browser infrastructure reliability, built and published by Halluminate in collaboration with Anchor in November 2025, with an open-source harness on GitHub so the runs can be reproduced. It comprises 292 tasks across 292 websites (83% commercial domains), each a read-based workflow of 5 to 25 steps, executed on the Browser Use agent framework and graded by humans with LLM-judge validation.
The benchmark's core metric is stealth failure rate: the share of agent trajectories that failed due to proxy and captcha issues. Lower means the infrastructure gets blocked less and the agent completes more work.
| Configuration | Stealth failure rate |
|---|---|
| Anchor Browser (stealth) | 1.7% |
| Hyperbrowser (stealth) | 2.7% |
| Browserbase (stealth) | 3.4% |
| Steel (no stealth) | 4.7% |
| Anchor Browser (no stealth) | 5.0% |
| Hyperbrowser (no stealth) | 6.5% |
| Steel (stealth) | 6.8% |
| Browserbase (no stealth) | 6.8% |
Halluminate's write-up: "Anchor Browser's stealth mode stood out as the most reliable, triggering blocks only 1.7% of the time." Anchor's engineering notes on the result, including the choices behind it (fast machines, real headful browsers, fine-grained user-like control, and Anchor's custom Chromium build), are published on the Anchor blog.
Browser Use stealth benchmark: 77%, ahead of every other infrastructure provider tested
In March 2026, Browser Use, the open-source agent framework, published a stealth benchmark built from its own production data: 300,000 security check events from live traffic, distilled to the 71 websites where agents most often get blocked, categorized by antibot vendor (Cloudflare 23 sites, PerimeterX 18, DataDome 13, Akamai 8, reCAPTCHA 6, others 3). Each provider ran the same three-step task per site (search, visit a page, read a link) with no authentication, and an LLM judge scored only whether the session was blocked.
| Provider | Pass rate |
|---|---|
| Browser Use Cloud (the benchmark's publisher) | 81% |
| Anchor Browser | 77% |
| Onkernel | 67% |
| Steel | 47% |
| Browserbase | 42% |
| Hyperbrowser | 40% |
Browser Use included its own cloud product in the test and it scored highest. Among the dedicated browser infrastructure providers, Anchor led at 77%, thirty points ahead of Steel, Browserbase, and Hyperbrowser. The authors note the honest caveat that applies to every benchmark in this category: proxy quality varies by the hour, sites update their configurations, and captcha difficulty fluctuates.
AIMultiple remote-browser benchmark: 82% composite, 91% features
AIMultiple, an independent research firm, benchmarks eight remote-browser services on live agent workloads: 160 automated tasks (four scenarios run five times per service: e-commerce purchase, lead generation via search, travel booking, web form submission) plus a load test with 250 concurrent agents. Last updated June 30, 2026.
| Provider | Composite | Task success | Speed | Features |
|---|---|---|---|---|
| Bright Data | 97% | 95% | 100% | 95% |
| BrowserAI | 87% | 85% | 90% | 86% |
| Anchor Browser | 82% | 70% | 86% | 91% |
| Steel | 72% | 70% | 99% | 45% |
| Browserbase | 65% | 50% | 94% | 50% |
| Hyperbrowser | 62% | 60% | 84% | 41% |
| ZenRows | 57% | 55% | 78% | 36% |
| Airtop | 44% | 40% | 42% | 50% |
In the full eight-vendor field, Bright Data, a proxy and web-data platform, posted the highest composite. Among the four providers that appear across all three benchmarks on this page (Anchor, Steel, Browserbase, Hyperbrowser), Anchor posted the highest composite (82%) and the highest features score (91%, second only to Bright Data's across the whole field). On task success, Anchor and Steel tied at 70%, ahead of Hyperbrowser (60%) and Browserbase (50%).
What these benchmarks measure
All three evaluations test the same underlying question from different angles: does an agent session survive detection and finish its work on real, protected websites? That is the reliability question for production automation, because antibot systems cannot distinguish an authorized agent from abuse, and every blocked session is a failed workflow. Anchor's approach to this problem, including stealth configuration and support for Cloudflare's Web Bot Auth standard for declared agents, is documented in the stealth docs.
These benchmarks do not measure everything a production evaluation weighs. Authentication depth, deployment options, compliance posture, and pricing behavior at volume are covered with sources on Anchor's comparison pages: Anchor vs. Steel and Anchor vs. Browserless, with session debugging covered in Debugging browser sessions in production.
Every vendor, Anchor included, is happiest sharing the benchmarks it wins. Two of the three above ship public repositories (BrowserBench, Browser Use benchmark), so the runs can be reproduced, and the test that matters most is your own workload. Anchor's free tier runs 5 concurrent browsers, enough to reproduce these benchmarks or run your own evaluation against the documentation's quickstart.