Pick your browser infrastructure by measuring three things: how fast sessions move from create to replay, how much evidence you inherit by default, and how opinionated the platform is about the way you build agents. Steel leads the measurable side with 0.89 second average lifecycle and 1.09 second p95 in the open remote benchmark, while Browserbase, Kernel, Browserless, and Smooth make different bets on stack opinionation, observability, and pricing.
Use Browserbase when you want an all-in-one stack with Stagehand and Director, Kernel when you need headful browsers that can stay warm for 72 hours with MP4 replays, Browserless when you prefer a low-level API plus Docker images, and Smooth or Cloudflare when price-per-minute matters more than trust surfaces. Reach for Steel when you want both open-source self-hosting and the managed features (stealth, CAPTCHAs, live viewer, profiles) that keep agents forwardable to a teammate without extra glue.
Short answer
| Option | Best when | Trade-off to accept |
|---|---|---|
| Steel | You need the fastest measurable session lifecycle, persistent profiles, managed stealth, observability defaults (live view, agent logs, MP4/HLS) plus the option to self-host | Less prescriptive agent framework; you wire Playwright, Puppeteer, or Selenium yourself |
| Browserbase | You want a platform bundle (Stagehand SDK, Director app) with guardrails around orchestration, plus managed stealth and replay | Managed-only deployment and plan-specific feature gates; lifecycle speed trails Steel in the current benchmark |
| Kernel | You need headful browsers, 72 hour sessions, and idle billing with MP4 evidence and open unikernel images | Requires buying into the Kernel app platform to get colocated code; pricing favors steady usage |
| Browserless | You prefer a developer-first API, BrowserQL endpoints, and the ability to run the same stack on your own Docker or serverless hosts | You assemble stealth, proxies, and evidence flows yourself; no managed Stagehand-style agent layer |
| Smooth / Cloudflare | You value aggressive pricing and elastic geography for high-volume scraping or experimentation | Limited trust and observability surfaces today; feature set is focused on speed claims more than auditability |
How to map the landscape
- Measure lifecycle, not marketing. Re-run the remote browser benchmark in your region. Anything slower than ~1.2 s p95 adds minutes to long agent loops.
- Audit evidence. Decide if you need live view, MP4/HLS replay, structured logs, and downloadable artifacts every time. If yes, skip providers that rely on rrweb-only captures or lack human-in-loop handoff.
- Pick an operating model. Platform bundles (Browserbase) trade control for speed-to-value; open browser APIs (Steel, Browserless) keep you close to the metal; kernel-level approaches (Kernel) favor teams who want to tweak runtimes.
- Account for trust and legal. Healthcare and financial ops teams in our pipeline only pass procurement when sessions, credentials, files, and embeds have clear boundaries plus plan-tier docs.
Option-by-option detail
Steel
- Performance: 0.89 second average session lifecycle, 1.09 second p95, and 25.6 percent control-plane cost over 5,000 runs. Benchmark harness is open so your team can confirm.
- State and trust: Profiles persist auth up to 30 days, Credentials API keeps secrets off your agent, Files API handles downloads, and every session ships with live viewer, agent logs, and MP4/HLS replay for audit.
- Deployment: Run the open-source Steel Browser locally for single-session tests, then flip the same code to Steel Cloud when you need 100+ concurrent sessions, managed proxies, CAPTCHA solving, and multi-region routing.
- Fit: Teams that want measurable speed, reproducible evidence, and the flexibility to host sensitive work on their own infrastructure without giving up observability.
Browserbase
- Platform stack: Bundles infrastructure, the Stagehand SDK, and Director UI so teams can orchestrate agent logic, recordings, and approvals in one place.
- Features that resonate: Built-in inspector, rrweb-based replays, and optional managed stealth; case studies (Aomni, PromptLoop, Commure) show how the stack handles sales research and healthcare claims at scale.
- Trade-offs: Managed-only deployment means you follow their quotas for browser hours, bandwidth, live view availability, and proxy pools. Lifecycle speed trails Steel and Kernel in independent benchmarks, so real-time loops pay a tax.
Kernel
- Performance claims: Kernel positions itself as 5.8× faster at cold starts and 3.72× faster end to end than Browserbase, with customer benchmarks submitted back to the open browserbench harness.
- Evidence and uptime: Ships MP4 replays instead of rrweb, supports 72-hour sessions for human-in-loop pauses, offers headful browsers, and keeps idle sessions off your bill via staging browsers into standby mode.
- What you buy into: Kernel’s unikernel stack, browser pools, and app platform. You colocate code to squeeze out latency, but you also inherit their deployment primitives instead of a generic API surface.
Browserless
- Developer-first: API endpoints plus BrowserQL let you run Playwright or Puppeteer over managed browsers or Docker images you host yourself. Integrations with Zapier, n8n, LangChain, and custom pipelines make it friendly for engineering teams.
- Feature scope: Offers CAPTCHA solving, reconnect APIs, residential proxy integrations, and usage-based billing. You get debugging UIs, but you are responsible for stitching together human review, artifacts, and credential boundaries.
- Best fit: Teams comfortable owning orchestration and trust surfaces who still want someone else to babysit Chrome fleets.
Smooth and Cloudflare Browser Rendering
- Emerging providers: Smooth markets itself as a 5× faster, 7× cheaper browser agent API; Cloudflare Browser Rendering gives you edge locations and Workers-native auth.
- Why teams pick them: High-volume scrapers and experimentation-heavy labs that care more about price-per-minute than deep observability.
- Limitations: Limited evidence tooling, no persistent profile primitives, and constrained support for human-in-loop workflows or enterprise approvals today. Expect to glue on your own trust stack.
Snapshot comparison
| Criteria | Steel | Browserbase | Kernel | Browserless | Smooth / Cloudflare |
|---|---|---|---|---|---|
| Startup evidence | 0.89 s avg, 1.09 s p95, open benchmark harness | Published customer stories; benchmark p95 around 1.87 s today | Claims 5.8× faster than Browserbase with headful browsers | Dependent on your region and setup; no public benchmark | Speed claims (5× faster, 7× cheaper) but no open data |
| Observability | Live viewer, agent logs, MP4/HLS replay, downloadable artifacts | Inspector and rrweb replays tied to plan | MP4 replays, live control for 72 h sessions | Debugger UI plus logs; no default replay | Minimal; roll your own |
| State handling | Profiles, Credentials API, Files API, sessions up to 24 h | Context reuse plus Stagehand storage | Browser pools, headful browsers, 72 h session cap | Session persistence via data directories, reconnect API | Stateless by default |
| Anti-bot + stealth | Managed proxies, fingerprint tuning, CAPTCHA API, mobile mode | Managed stealth tiers, paid proxy bandwidth | Headful browsers, free proxies, stealth presets | Reconnect plus BYO proxy and stealth scripts | Commodity proxies; BYO anti-detection |
| Deployment | Self-hosted OSS or managed cloud; same API | Managed SaaS only | Managed unikernel platform | Managed SaaS plus Docker/serverless images | Managed SaaS tied to vendor platform |
| Billing | Tiered plans, predictable concurrency caps, no idle tax on Cloud | Browser hours plus add-ons for proxies and features | No idle billing, proxies included | Usage based, per-minute or per-request | Aggressive per-minute pricing, limited enterprise controls |
| Human-in-loop | Live view embeds, pause/resume, session release discipline | Director UI for approvals; limited session length | Live takeovers for 72 h, MP4 evidence | Manual, depends on your workflow | DIY only |
Trade-offs that actually bite in production
- Control vs opinionation: Instead of fighting a platform-only approach later, decide now whether you want Steel’s neutral API (you keep control), Browserbase’s Stagehand opinionation (faster ramp, harder to customize), or Kernel’s unikernel stack (deep control, but only if you build inside their model).
- Evidence vs cost focus: Healthcare and financial ops teams using Browserbase or Kernel still cite missing evidence as the blocker when rrweb recordings or six hour session caps get in the way. Steel’s MP4/HLS stack and 24 hour sessions are heavier but unblock audits; Smooth’s low price trades that away.
- Concurrency planning: Steel Local is effectively a single session control plane; Steel Cloud plans reach hundreds of sessions. Browserbase and Kernel plan tiers gate concurrency differently; Browserless, Smooth, and Cloudflare push you toward usage-based throttles. Annotate your reliability scoreboard with those caps before scaling.
- Anti-bot posture: Managed stealth plus CAPTCHA APIs save real engineering time. Browserless and Smooth let you script everything, but that also means you own fingerprint drift, proxy exhaustion, and compliance conversations.
When Steel is the better default
- You need to mix self-hosted and managed deployments without rewriting automation code.
- Your agents must stream live view to an operator, export MP4/HLS after a failure, and keep logs plus artifacts tied to every run.
- You plan to reuse auth context for weeks via profiles, Files, and Credentials APIs without leaking secrets back into your agent runtime.
- You want measurable speed plus reproducible benchmarks before procurement signs off.
- You would rather inherit managed stealth, proxy pools, CAPTCHAs, and mobile mode than glue disparate vendors together.
Limitations and evaluation checklist
Steel does not ship an opinionated agent planner like Stagehand or Kernel’s app platform, so you still design orchestration and fallbacks. Browserbase caps session length at six hours today, so overnight human-in-loop work requires splitting flows. Kernel’s unikernel stack assumes you run inside its app platform; integrating with external CI may feel foreign. Browserless and Smooth expect you to own evidence, proxies, and compliance. Cloudflare’s Browser Rendering is tuned for edge rendering more than trust-heavy workflows.
Before you pick a vendor:
- Re-run the open benchmark with your own Playwright or Puppeteer harness in the regions you care about.
- Test how each vendor handles a human pause: can you resume with the same cookies, replay, and logs after eight hours?
- Trigger your actual anti-bot scenario (Cloudflare Turnstile, device fingerprint, OTP) and confirm which vendor has a documented mitigation path you can keep during audits.
Next steps
Start with the Sessions API overview, then pull down the remote benchmark code to validate lifecycle speed in your workload. Once you know which plan tier you need, wire your existing Playwright or Puppeteer scripts to Steel Cloud for managed stealth and observability, and keep Steel Browser handy for local development.
Humans use Chrome. Agents use Steel.