Cloud browser or local browser for AI agents: 5 costs compared

A cloud browser is one API call away, which is why most agent projects start there. The surprises come later: a bill with three meters on it, and a session that ends in the middle of a task. We read the pricing and limits pages of 5 vendors on 26 September 2026 and set them against a browser that runs on your own machine.

What is a cloud browser for AI agents?

A cloud browser for AI agents is a Chromium instance in a vendor's data centre that your agent drives over a CDP address and pays for by the hour. Five vendors sell it under that name today: Browserbase, browser-use Cloud, Steel, Kernel and Hyperbrowser. Your agent code stays with you. The browser, its profile and usually its exit address live with the vendor.

Which 5 costs differ between cloud and local?

Five costs separate the two: how long one session may run, what a browser hour costs, what a gigabyte of proxy traffic costs, whether the browser renders on a real GPU, and who decides the fingerprint. The table uses each vendor's entry plan that allows ongoing work; an empty cell reads "not published" because we found no number in the vendor's own pages.

Two of the five never show up on an invoice: the GPU and the fingerprint. Kernel's own documentation says that software-rendered browsers produce pixel hashes that match no real consumer GPU, "which is a strong bot signal on sites with rendering-based fingerprinting." Browser Use's stealth documentation says canvas, WebGL, fonts and navigator are randomised per session. We measured what randomising a fingerprint costs when we tried it ourselves.

We took the noise out again. Canvas, WebGL, WebGPU, text measurement and audio now report the device's own values, so a checker compares the picture with a real GPU and finds one.

What does one agent cost per month?

One agent that works 8 hours a day on 22 working days uses 176 browser hours; with 5 GB of proxy traffic that comes to between 28.52 USD and 656.88 USD a month in the cloud, depending on the vendor and on whether it renders on a GPU. The figures below are calculated from the prices in the table above, not billed.

The session caps change the work, not only the bill. Browserbase's 6 hours split the 8-hour day into two sessions, and Steel Launch's 15 minutes split it into 32. Every new session has to find its way back into a signed-in state before the agent can continue.

When is a cloud browser the right choice?

A cloud browser is the right choice when there is no machine you can keep running, or when a buyer requires the vendor's compliance certificates. It is also the quickest way to try an agent for an afternoon. The trade shows up once the agent has to keep an account: a local browser for AI agents keeps the same profile, the same logins and the same device from one run to the next, because nothing on it is reset between sessions. How a computer use agent and a browser agent each see a page is covered in computer use agent or browser agent.

What changes when the browser runs on your machine?

Three things change when the browser runs on your machine: it renders on the real GPU, its traffic leaves through the exit you choose, and the number of profiles is not metered. A Scalebrowser lease runs 900 seconds and is renewed while the agent works, so no session has a fixed end (concepts). The costs move elsewhere: you keep the machine on, and you buy proxy traffic from a provider yourself. How a run is recorded as one still image per browser step is covered in AI agent observability.

How does browser-use Cloud compare?

browser-use Cloud is the cheapest cloud line in our calculation, 28.52 USD for the example month, because it charges 0.02 USD per browser hour. Its fingerprint is the difference: it is randomised per session by design, which is the same kind of change our canvas bench flagged in 5 of 5 profiles. The open-source browser-use library itself can also drive a local browser. Vercel's agent-browser is compared with a browser that has an identity in Vercel agent-browser vs a browser with an identity.

So which should you pick?

Pick by where the agent's accounts live and how long its work runs.

If your agent already runs on a machine that stays on, start there, and keep the cloud for the runs that have nowhere else to go.