How much time does the browser add to a task?
A session creation number measures one API operation. A usable browser also needs a connection and a navigation, and a long-running workflow pays for many actions after startup. Compare lifecycle latency and running-task throughput separately.
A workflow with one navigation and a hundred browser actions may favor a different execution path than a workload that opens and closes a browser for every action. Test both shapes before making a platform-wide speed claim.
Compare the browser lifecycle
Notte: Create an owned browser session, connect Playwright or Puppeteer over CDP, and stop the session when the work finishes. Browser lifecycle management remains necessary even when the control socket disconnects.
Steel: Steel sessions expose browser connection details for existing automation code. Create and release the owned session through Steel's lifecycle API; a browser library is the controller, not the infrastructure scheduler.
Compare the evidence from a run
Notte: Live view shows an active browser and can be embedded in an application. Recordings let you review a finished session. Function runs have their own history and results, so keep the run ID alongside the browser session ID.
Steel: Steel documents live and past sessions as well as Agent Traces. Browser recordings answer what happened on the page; traces add context around the agent or application driving it.
Compare deployment ownership
Notte: Use a remote browser from your existing process or deploy the browser automation as a Notte Function. These choices put lifecycle, scheduling and monitoring responsibilities in different places.
Steel: Compare managed Steel with the documented self-hosted path explicitly. Self-hosting gives the team infrastructure responsibility that a hosted browser subscription would otherwise cover.
How to evaluate this choice
For short tools, lifecycle overhead can dominate. For a browser that stays open, action latency, task completion and drift matter more. Select the measurement that resembles your application.
- Use the same target page, browser configuration and controller workload.
- Record the runner and browser regions, account plan and concurrency.
- Keep the measured sample count and any warmup or retry behavior visible.
- Calculate tail percentiles from per-session totals, not sums of phase percentiles.
- Classify configuration failures before interpreting them as product failures.
Published browser measurements
These are the latest provider runs in the saved BrowserArena snapshot checked on October 18, 2026. Dates and runner environments can differ. A lifecycle test opens, connects, navigates and releases a browser; the throughput test runs a separate browser action sequence. Neither measures managed authentication or deployed Function reliability.
Use these observations to choose what to test in your application. An account-configuration failure, an incompatible page action and a timed-out navigation require different interpretations. BrowserArena is maintained by Notte and publishes its source and methodology.
| Provider | Lifecycle median | Lifecycle P95 | Actions / second | Run dates |
|---|---|---|---|---|
| Notte | 332 ms | 451 ms | 3.14 | 2026-10-18 / 2026-10-18 |
| Steel | 1496 ms | 1950 ms | 3.54 | 2026-10-18 / 2026-10-17 |
Sources: browserarena.ai [1] ↗browserarena.ai [2] ↗github.com [3] ↗
Which approach fits?
Choose Notte when your browser workflow needs saved logins, credential fill or scheduled Python code alongside the browser.
Notte tradeoff: A saved profile restores browser state; it does not prove that a website still considers the user signed in. Use Managed Auth when the workflow needs connection readiness before it starts.
Choose Steel when you want browser infrastructure that can fit a self-hosted deployment, with profiles and credentials available through browser-focused APIs.
Steel tradeoff: Hosted and self-hosted Steel are different operating choices. A cloud feature's documentation does not establish identical packaging, retention or operational support in a self-hosted installation.
A useful acceptance check
Write a pass/fail result for the real task before running the evaluation. Record the account, configuration and code version next to that result. Repeat failures need an explanation, not an automatically different browser configuration.