Anchor sells enterprise via a Talk-to-Expert form. humanbrowser lets you try it for $1, with a public price list, and an open-source core on GitHub. Same job — different go-to-market.
| Dimension | humanbrowser | Anchor Browser |
|---|---|---|
| Buying motion | Self-serve · try it for $1 | $0 / 5 credits + Talk to Expert |
| Per-line-item pricing | Public — minutes, GB, solves, tokens | Credit pack, then sales call |
| Open-source core | @virixlabs/humanbrowser on npm | Closed source |
| Execution model | CUA + DOM hybrid · auto-routed per site | DOM / CDP only |
| Public benchmark | 6 engines × 12 LLMs · OSS | None published |
| A2A (v0.3) endpoint | Native · agent card discoverable | Not supported |
| Cloudflare Web Bot Auth | Roadmap (A2A identity already exposed) | Launch partner · live today |
| Live viewer + human takeover | MJPEG + one-click HITL · free | Live view · takeover gated |
Anchor Browser is a New-York-based cloud-browser company, $6M seed in early 2026, founded by Unit 8200 alumni, and one of the launch partners for Cloudflare's Web Bot Auth program. The product is real and the team is strong. The motion is enterprise: a $0 free tier capped at 5 credits, then a Talk-to-Expert form for anything past that, and the named-account playbook Cloudflare partnerships imply.
humanbrowser is the opposite shape on every axis except the underlying engineering problem. Same job — give an AI agent a cloud browser that can execute user-authorized tasks on real websites with a clean residential connection. Different go-to-market: self-serve, public pricing, open-source core, public benchmarks.
None of that is criticism of Anchor's strategy — enterprise sales is a legitimate playbook for infrastructure with a high integration cost. It is criticism of misallocation. If you are an AI-agent shop, a YC team, or a developer who wants to ship something this afternoon, the enterprise-sales motion is friction you should not have to swallow.
$1, no demo call. Top up from $5, balance never expires.agent.humanbrowser.cloud/a2a plus discoverable agent card at /.well-known/agent-card.json. Claude Code, Codex, LangGraph, CrewAI, OpenAI Agents SDK, Google ADK all plug in straight.google/gemini-2.5-flash-lite at $0.023/agentic task as the cost-leader.Most cloud-browser products execute through DOM selectors only — find element, click element, type into element. That works for 80% of websites. The other 20% — heavy canvas surfaces, drag-and-drop interfaces, complex shadow DOM, pages that actively detect synthetic CDP input — is exactly where AI agents fall over.
humanbrowser ships a real Computer-Use-Agent backend (xdotool + scrot driving Xvfb, fronted by our in-house Computer-Use model) for those targets. Our public MVP test on example.com landed at 25.8s end-to-end, 5 steps, $0.044 cost. Because the CUA path uses real human-input primitives (no CDP, no JS hooks injected into the page) it is also resilient to the WebDriver-style detection that trips up DOM-only stacks. WebDriver=missing/passed on bot.sannysoft.com on the same MVP run.
Per-domain rules in our knowledge base decide automatically whether a site goes through CUA or the cheap DOM path. You do not pick the backend — the system does, based on what worked last time.
| Line item | humanbrowser | Anchor Browser |
|---|---|---|
| Entry | $1 first top-up · about 4 typical tasks | $0 / 5 credits |
| Subscription | None · pay-as-you-go | Not publicly disclosed past free tier |
| Browser time | $0.10/hr | Talk to Expert |
| Residential connection | $4/GB · 75 countries | Talk to Expert |
| Interactive challenge solving | $0.005/solve · billed on success | Talk to Expert |
| AI planning | $0.02 / agent step, AI included | Talk to Expert |
| Per-task economics (median) | $0.24 median, all-in (about 8 agent steps) | Not publicly disclosed |
| Per-task economics (cached) | $0.03–$0.05 | Not publicly disclosed |
| Balance expiry | Never | Credits expire on free tier |
This is not a gotcha. Enterprise vendors hide pricing because the deal shape varies per account — multi-year commits, custom SLAs, dedicated capacity. That math makes sense for large procurement. It does not make sense for a developer or a 4-person AI shop that wants to budget a project in an afternoon. The simple answer is: pick the vendor whose pricing-disclosure posture matches your stage.
google/gemini-2.5-flash-lite on Class-C agentic tasks.Anchor publishes no equivalent benchmark. You take the marketing claims on trust, or you build your own test rig and run both. We did the second step in public so you do not have to.
This page exists to help you make a real decision, not to win every comparison. Anchor wins these:
If you already have an Anchor-driven script, migration is one integration swap. POST your task spec to https://agent.humanbrowser.cloud/a2a — navigate, click, extract, log in, export. humanbrowser plans and executes it inside a session with a clean residential connection, the right execution backend (CUA or DOM, auto-routed), and the live viewer attached. You stop wiring proxy configs separately, you stop calling fallback solvers as a side-channel. Try it for $1, smoke-test one representative flow, then cut your highest-friction script over first.
import { Anchor } from '@anchorbrowser/sdk'; const anchor = new Anchor({ apiKey: process.env.ANCHOR_KEY }); const session = await anchor.sessions.create({ // Pricing per minute / GB / solve negotiated // after Talk-to-Expert conversation. region: 'us-east' }); const page = await session.openPage('https://target.com/login'); await page.type('#email', process.env.USER_EMAIL); await page.type('#password', process.env.USER_PASS); await page.click('button[type=submit]'); // ...your DOM-by-DOM logic...
const res = await fetch('https://agent.humanbrowser.cloud/a2a', { method: 'POST', headers: { 'Content-Type': 'application/json', 'Authorization': 'Bearer ' + process.env.HB_TOKEN }, body: JSON.stringify({ jsonrpc: '2.0', id: 'login-1', method: 'tasks/send', params: { skill: 'login_and_extract', message: 'Log in to target.com and export CSV', country: 'gb', profile: 'reports-bot' } }) }); const { result } = await res.json(); // Residential connection, auto-routed backend, // live viewer attached, result.artifacts has the CSV.
Left side: a DOM-by-DOM script you maintain forever. Right side: declare the goal, hand off the execution. When the site updates and a selector breaks, the A2A version re-plans; the DOM version pages you at 3am.
Try it for $1 — about 4 typical tasks; top up from $5. $0.10/hr browser, $4/GB residential, $0.005/solve, $0.02 per agent step with the AI included. Open-source runner on GitHub. Balance never expires.
Try it for $1Yes, for the developer-self-serve use case. Anchor Browser is an enterprise-sales product — $0 / 5 credits free tier plus a Talk-to-Expert path for anything beyond. humanbrowser is the opposite shape: try it for $1 with no sales call, public per-line-item pricing, open-source core on npm as @virixlabs/humanbrowser, and an A2A (v0.3) endpoint discoverable at /.well-known/agent-card.json. If your buyer is a Fortune 500 procurement officer, Anchor's relationships and partnerships fit better. If you are a developer or AI-agent shop that wants to start the same hour and see the bill before signing anything, humanbrowser is the cleaner fit.
Anchor publishes a $0 / 5 credits trial and routes everything else to Talk to an Expert. humanbrowser is fully posted: $0.10/hr browser time, $4/GB residential connection, $0.005/solve for interactive challenges, $0.02 per agent step with the AI included. On our production the median task took about 8 agent steps and cost $0.24 all-in (132 customer tasks, Sep 2026). Our internal LLM benchmark (12 sessions, ~$4 total spend) found google/gemini-2.5-flash-lite the cost-leader at $0.023 per Class-C agentic task. No subscription, no minimum, balance never expires.
Yes. humanbrowser exposes a plain HTTP A2A (v0.3) endpoint at agent.humanbrowser.cloud/a2a plus a discoverable agent card at /.well-known/agent-card.json. Any A2A-speaking agent — Claude Code, Codex, LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, OpenClaw — reads the card and sends tasks. Scripts written against Anchor's session API are wrapped as A2A tasks and dispatched. The open-source npm package @virixlabs/humanbrowser also gives you a local-first execution path if you want to keep the agent on your machine and only call the cloud for the browser session.
Cloudflare's Web Bot Auth — a cryptographic agent passport launched May 2026, with Anchor Browser as one of the announcement partners — is on humanbrowser's near-term roadmap. Our A2A endpoint already exposes an agent identity, so wiring a signed-key handshake is incremental. For sites that have opted in to Web Bot Auth verification, this is the cleanest path; for the long tail of sites that have not (the overwhelming majority today), a clean residential connection plus a real browser fingerprint and human-paced interaction is what actually works. If your specific enterprise contract requires the Cloudflare partner-program logo on day one, Anchor is the safer procurement choice this quarter.
Because nobody else in this category does, and engineering teams making a buying decision deserve real numbers. We run 6 cloud-browser engines × 12 LLMs against representative sites and publish the full methodology and raw data at github.com/VirixLabs/stealth-benchmark. No vendor scoring its own homework, no hidden test set. If you find a flaw in the test design, file an issue and we will fix it in the next run. That public-data posture is the inverse of how enterprise-sales vendors typically operate, and we think it is the correct default for infrastructure that AI agents are betting on.
Most cloud-browser products execute through DOM selectors only — find element, click element, type into element. humanbrowser ships a Computer-Use-Agent path (xdotool + scrot on Xvfb, driven by an in-house Computer-Use model) for sites where DOM is the wrong layer — heavy canvas, drag-and-drop, complex shadow DOM, or pages that actively detect synthetic CDP input. Per-domain rules in our knowledge base auto-route the harder targets to CUA and keep the cheap-and-fast DOM path for the rest. Our public MVP test recorded a 25.8s end-to-end task on example.com at $0.044 cost — no CDP, no JS hooks, just a real human-input pipeline.
Yes. The runner is published on npm as @virixlabs/humanbrowser (Apache-2.0). You can run a full local agent loop offline, or point it at the hosted cloud (humanbrowser.cloud) when you need shared residential connections, the live viewer, or A2A discovery. The stealth-engine benchmark code is also OSS at github.com/VirixLabs/stealth-benchmark. Compare to Anchor, which is closed-source by design.
If you need the Cloudflare Web Bot Auth ecosystem integration as a contractual deliverable for an enterprise deal that is being signed this quarter, or you are already inside the Unit 8200 / Israeli-cybersecurity network deal flow, Anchor's relationships and brand recognition beat ours on that specific path. Choose them then. For developer self-serve, transparent pricing, open-source core, public benchmarks, and a CUA + DOM hybrid you can switch on in one config flag — humanbrowser is the cleaner fit.
Loading secure checkout…