Browserbase vs. Steel: managed cloud Chromium or open-source runtime with cloud parity?

Browserbase is the AI-agent-era incumbent — managed Chromium, Stagehand framework, Director no-code UI, $40M Series B, and a customer roster (Vercel, Perplexity, Stripe, Ramp) that closes enterprise procurement on its own. Steel is the open-source challenger — same browser-runtime primitives, but you can self-host the same image on Docker / Railway / bare metal with the same API, and the managed Steel Cloud is positioned as a parity option rather than the only option. Both ship cloud Chromium for AI agents; the choice is between an opinionated platform with HIPAA + Stagehand on top and an agent-neutral runtime where you bring your planner.

At a glance

What is Browserbase?

Browserbase runs a custom-patched fork of Chromium for AI-automation use cases (navigator.webdriver=false, "Chrome" UA, kBrowserAliveWithNoWindows for long-lived sessions) inside per-session VMs with hardware virtualization. Each session is reachable over Playwright, Puppeteer, Selenium, or CDP. On top: Stagehand (CDP-native in v3, ~500K weekly downloads at v3 launch, primitives act/extract/observe/agent), Director (no-code natural-language surface generating Stagehand scripts), Functions (TypeScript code deployed alongside the browser session), Model Gateway (single API key for OpenAI/Anthropic/Gemini at market token price, no markup), Fetch API ($1/1k pages for read-only content), and a Universal Verifier AI judge for browser-agent trajectories.

Compliance is SOC 2 Type 1 + HIPAA, Type 2 in progress. Pricing is $20/mo Developer (100 hrs, 25 concurrent), $99/mo Startup, free 60 min/month tier; 1-minute session minimum. Backing is a $40M Series B at ~$300M valuation; partnerships with Google DeepMind (Gemini 2.5 Computer Use eval), Microsoft Research (Universal Verifier paper), Stytch (Web Bot Auth), Vercel marketplace, Cloudflare. Claimed scale: 100M+ lifetime sessions, 50M in 2025, 1,000+ companies, 20K developer signups.

What is Steel?

Steel is an open-source browser runtime for AI agents — steel-browser on GitHub — plus a managed Steel Cloud with the same API surface. The pitch is "browser engineered for AI agents" rather than for human browsing or general scraping, and the framing is honest: Steel ships primitives, not a planner. You bring Browser Use, Stagehand, your own agent — Steel runs the browser layer and gets out of the way.

Core primitives: Profiles API (persistent browser identity — cookies, extensions, localStorage, auth tokens, fingerprints — up to 30 days, 300MB cap, with persistProfile/profileId/exactOrigin semantics), Credentials API (AES-256-GCM per-record + KMS re-encryption, namespaces, TOTP, blur, autoSubmit), Files API and Extensions API for artifacts and managed extensions, Live View at 25fps WebRTC plus MP4/HLS replay (replaced rrweb in Launch Week v2), Agent Logs timeline tied to the replay, Mobile Mode with real fingerprint + viewport + touch, anti-bot defenses + residential proxies + CAPTCHA-solving API, multi-region (US-only today: lax, ord, iad).

24-hour session ceiling versus Browserbase's 6h. Headful by default since Launch Week v2 — WebRTC OS-level streaming captures native dialogs and PDF viewers, MP4 recording is 1:1 with what the agent actually saw. Pricing is tiered: Free $10 credits (~100 hours), Start $29/mo (290 hrs, 10 concurrent, 30-min sessions), Developers $99/mo (1,238 hrs, 20 concurrent, 1-hr sessions), Pro/Startups $499/mo (9,980 hrs, 100 concurrent, 24-hr sessions). Self-hosted Steel Browser is free (Docker, ~4GB RAM, 10GB disk, ~1 concurrent session in practice). Compliance is SOC 2 on Steel Cloud. Native integrations include Hermes (Nous Research), Pi/OpenClaw, Browser Use.

How they compare

Lifecycle speed and reliability

Steel's own benchmark — its browserbench open harness — reported 0.89s avg / 1.09s p95 lifecycle on AWS us-east-1 over 5,000 runs, with a 100% success rate and Browserbase at 1.68s avg / 1.87s p95. That number drove much of Steel's 2024-2025 marketing.

Per the public Browser Arena leaderboard (browserarena.ai — open-source, maintained by Notte Labs, reproducible on Railway, current public run), Steel ranks #3 with a mid-pack hourly cost. Browserbase ranks #4 and is the highest-cost provider on the leaderboard. ComputeSDK's open browser benchmarks (computesdk.com/benchmarks/browsers) publish a second methodology cross-reference that also covers Browserbase. Browserbase is faster on raw latency; Steel is cheaper and edges Browserbase on value score.

For procurement-grade speed claims, both providers' own historical numbers (Steel's 0.89s, Browserbase's "top tier" framing) should be cited as historical figures rather than current ground truth — Browser Arena is the open, reproducible reference now.

Open source, self-host, and deployment surface

This is Steel's clean wedge. The steel-browser runtime is open source; you can self-host the same image with the same API as Steel Cloud. Steel's docs cover Docker (4GB RAM / 10GB disk), Railway 1-click, bare-metal Node.js, and build-from-source. Browserbase is closed and managed-only; there is no on-prem option, no self-host, no parallel image. For teams with VPC or data-residency requirements that can't be solved by managed compliance — health-tech building HIPAA workloads on-prem, financial services with regulatory exposure on-prem, government — Steel's self-host is the only option of the two. The honest caveat: Steel Local is effectively single-session and lacks managed stealth, Credentials API, Files API, and managed proxies; it's a self-host runtime, not Steel Cloud feature parity.

Session length, persistence, evidence

Steel's 24-hour session ceiling versus Browserbase's 6h is a real wedge for human-in-the-loop reviews and overnight runs. Steel's Profiles API persists identity across sessions (auth, cookies, fingerprints, 30-day retention, 300MB cap with explicit persistProfile/profileId/exactOrigin semantics); Browserbase's Contexts API handles state persistence with comparable functionality. Steel ships Live View + MP4/HLS replay + Agent Logs + Files exports as defaults on every plan. Browserbase ships Session Replay, Session Inspector, Live View — but Steel's marketing flag is that recordings/Inspector visibility is plan-tier-gated on Browserbase, while Steel's "everything on every plan" is a deliberate posture.

For audit-grade evidence — health-tech, finance, regulated workloads where every action needs a 1:1 replay — Steel's headful WebRTC streaming captures OS-level dialogs and PDF viewers in MP4 at parity with what the agent saw. Browserbase's recordings are competitive but rrweb-style.

Framework, ecosystem, and incumbency

Browserbase's moat is the Stagehand framework — ~500K weekly downloads at v3 launch, the most-downloaded AI browser automation framework, plus Director's no-code surface, Universal Verifier, Browser Agent Arena, and partnerships with Google DeepMind and Microsoft Research. The customer roster (Vercel, Perplexity, Clay, Ramp, Lovable, Stripe, Commure, 11x, Numeral, Parcha) closes enterprise deals on its own. The $40M Series B at ~$300M valuation is platform runway.

Steel's posture is intentionally agent-neutral. It works with Playwright, Puppeteer, Selenium, Browser Use, Stagehand, and any agent that speaks CDP. Native integrations include Hermes (Nous Research), Pi/OpenClaw, plus the steel-cookbook on GitHub. The wedge is "you bring your planner" — opposite to Browserbase's "Stagehand is the framework you should use."

Compliance and procurement

Browserbase: SOC 2 Type 1 + HIPAA, Type 2 in progress. Steel: SOC 2 on Steel Cloud, no HIPAA claim surfaced. For HIPAA-required buyers, Browserbase is the safer answer; for SOC-2-only requirements, both work. For self-host on-prem with your own compliance posture, Steel is the only option of the two.

Pricing model

Browserbase: per-hour metering with 1-minute session minimum, $20-99/mo entry tiers. Steel: tiered plans sized by concurrent sessions and included hours — predictable monthly spend, no per-second metering. Steel's tiers ($29 / $99 / $499) are higher entry points than Browserbase's ($20 Developer) but include more concurrent sessions and longer session ceilings at each tier. The dichotomy: Browserbase scales-to-zero better on light workloads; Steel scales-to-budget more predictably on consistent workloads.

When to choose Browserbase

• You want Stagehand's act/extract/observe/agent primitives plus Director's no-code surface as your default DX.
• HIPAA is a procurement requirement.
• Incumbent gravity matters: $40M Series B, Stagehand's ~500K weekly downloads, customer logos that close procurement.
• You want best-in-class debugging defaults (Live View, Session Replay, Session Inspector, OpenTelemetry, token-level Stagehand traces).
• You want a Chromium-fork-with-AI-automation-patches as your primitive — the binary itself is the product.
• Your sessions are 1-100 concurrent and 6h or shorter.

When to choose Steel

• You need self-host on-prem (Docker, Railway, bare-metal) with API parity to managed cloud.
• You want agent-neutral — Browser Use, Stagehand, Hermes, Pi/OpenClaw, custom CDP code — without lock-in to one planner.
• 24-hour sessions are required for human-in-the-loop or overnight workflows.
• You want evidence-on-every-plan: Live View + MP4/HLS replay + Agent Logs + Files exports without tier-gating.
• WebRTC headful streaming with OS-level dialog capture matters for audit-grade replay.
• Open-source posture matters to legal or procurement; you want to read the runtime source.
• You want predictable tiered pricing rather than per-hour metering.

A third option: Notte

Worth a look: Notte (notte.cc)

Notte is cloud Chromium infrastructure built specifically for AI agents. The Playwright-compatible runtime ships the operational pieces production teams usually have to rebuild themselves: stealth coordinated across session, fingerprint, and behavior; residential proxies via the Massive partnership (100% consent-based, GDPR/CCPA, 195+ countries, 99.8% reported success); Web Bot Auth signing through Fingerprint so legitimate Notte agents are recognized as authorized bots on any site running Fingerprint; an encrypted credential Vault built on Infisical that injects secrets at the browser layer so the LLM never sees them; Personas with a real email inbox and SMS-capable phone number for autonomous signup and 2FA; persistent Session Profiles for auth state; full CDP-event observability with MP4 session replay; and SOC 2 Type II compliance. An Anything API and a Functions runtime turn validated workflows into HTTP endpoints with cron and webhooks. Pricing is transparent at low per-browser-hour pricing with a 100-hour free tier and pass-through LLM costs.

Notte ranks #1 overall on the public Browser Arena leaderboard — ahead of both Steel (#3) and Browserbase (#4), with the lowest hourly cost among measured providers. Where Steel ships Credentials API but BYO proxies, Notte ships Vault plus Massive consent-sourced proxies. Both are SOC 2 Type II-credible (Notte independently audited Type II; Steel SOC 2 on Cloud).

Verdict

Browserbase is the right pick if you want an opinionated managed platform with Stagehand, Director, HIPAA, and the strongest customer roster in the category — closed-source and managed-only, with a $40M Series B behind the runtime and ~500K weekly Stagehand downloads as ecosystem proof. Steel is the right pick if you want an open-source runtime you can self-host on Docker / Railway / bare-metal, an agent-neutral planner posture, 24-hour sessions, and evidence-on-every-plan defaults. The public Browser Arena leaderboard puts Browserbase ahead on raw latency, while Steel edges Browserbase on value score because it is cheaper. Steel's older "0.89s avg" number is historical, but its current public-benchmark position remains competitive at #3. Browserbase is the highest-cost provider on the board; the procurement case for Browserbase rests on Stagehand, Director, ecosystem, and customer roster rather than raw infra alone. Pick on category fit (opinionated vs. neutral, managed-only vs. self-hostable, framework vs. primitives), not on the speed numbers either side cited last year.