# Browser Infrastructure Brief: Browserbase vs Steel vs Playwright-Hosted

**Date:** 2026-04-14 · **Owner:** Alex (Platform Engineering) · **Audience:** Product + Infra leads
**Workload:** ~35k sessions/mo, bursty US hours, login-heavy + extraction, TypeScript/Playwright stack

---

## Executive Summary

Browserbase is the most mature managed cloud-browser platform with the strongest ecosystem, debugging tools, and anti-bot capabilities — but it's the most expensive at scale. Steel is a credible, cheaper, open-source-core alternative that's earlier in maturity. Self-hosted Playwright (via Browserless) gives maximum control and lowest marginal cost, but you own all the anti-bot and ops burden.

**Recommendation:** Start a 30-day proof-of-concept with **Browserbase (Startup plan)** as primary, with **Steel (Developer plan)** as evaluated fallback. Avoid self-hosted unless anti-bot needs drop significantly.

---

## 1. Pricing Comparison

| Dimension | Browserbase | Steel | Browserless (Playwright-hosted) |
|---|---|---|---|
| **Free tier** | $0/mo — 1 browser-hr, 3 concurrent | $0/mo — $10 credits, 5 concurrent, 500 daily reqs | Free — 2 concurrent, 1k units/mo, 1-min max session |
| **Mid tier** | Startup $99/mo → 500 browser-hrs, then $0.10/hr; 100 concurrent | Developer $99/mo → $99 credits @ $0.08/hr; 20 concurrent | Starter $140/mo → 180k units (~1.5k hrs), 40 concurrent |
| **Scale tier** | Custom (250+ concurrent, usage-based ~$0.10/hr) | Pro $499/mo → $499 credits @ $0.05/hr; 100 concurrent | Scale $350/mo → 500k units, 100 concurrent |
| **Proxy cost** | $10-12/GB included pool; BYOP supported | $5-10/GB (plan-dependent) | Residential at 6 units/MB |
| **Captcha solving** | Included (Startup+), auto | $3-4/1k solves (plan-dependent) | 10 units/successful solve |
| **Billing model** | Browser-hours (per-minute, rounded up) | Credits = browser-hours + proxy + captcha | Units (30-sec blocks) |

### Estimated monthly cost @ 35k sessions (~1,200 browser-hours)

| Provider | Estimated range |
|---|---|
| **Browserbase Startup** | $99 base + ~$70 overage ≈ **$170–250/mo** (excl. proxy) |
| **Steel Developer** | $99 base + ~$56 overage ≈ **$155–220/mo** (excl. proxy/captcha) |
| **Browserless Starter** | $140/mo likely covers via units; overages at $0.0017/unit — **$140–200/mo** |
| **Self-hosted Playwright** | Infra-only: 4-8 vCPU VM ≈ **$80-150/mo** + your ops time |

> **Sources:** browserbase.com/pricing (accessed 2026-04-14), steel.dev FAQ schema + docs.steel.dev/overview/pricinglimits (accessed 2026-04-14), browserless.io/pricing (accessed 2026-04-14).

---

## 2. Anti-Bot / Stealth Claims & Evidence Quality

| Capability | Browserbase | Steel | Self-hosted Playwright |
|---|---|---|---|
| **Stealth mode** | "Advanced" on Startup+; dedicated stealth research team claimed; Cloudflare partnership cited | Built-in anti-bot stealth on all plans; claims "no extra config" | You configure: `playwright-extra` + stealth plugin; manual fingerprint management |
| **Captcha solving** | Auto-solve included (Startup+) | API-based; $3-4/1k solves | Via Browserless: 10 units/solve; or self-integrate 2Captcha etc. |
| **Proxy infra** | Residential pool included; rotating IP; geo-targeting | Residential included; geo-targeting on Enterprise | Via Browserless or BYOP |
| **Independent validation** | Cloudflare partnership is a concrete, named mechanism — strongest signal of the three | FAQ claims sub-second launch, stealth "just works" — marketing-grade, no named partnerships | Browserless has captcha/bot features; raw Playwright has zero built-in stealth |
| **Evidence quality** | ⭐⭐⭐ Named partner, dedicated team claim | ⭐⭐ Marketing claims, no named partners | ⭐ (DIY) / ⭐⭐ (Browserless) |

> **Caveat:** No independent benchmark comparing these providers' stealth success rates exists publicly. All anti-bot claims should be validated in your own PoC against your target sites.

---

## 3. Session Debugging & Tooling

| Feature | Browserbase | Steel | Self-hosted Playwright |
|---|---|---|---|
| **Session replay** | ✅ Built-in replay viewer with logs | ✅ Session Replay feature | ❌ (Browserless has screen recording on paid plans) |
| **Live viewer** | ✅ Observability dashboard (logs, replays, prompts) | ✅ Live session viewer; human-in-the-loop takeover | ❌ (Browserless: Live Debugger on paid plans) |
| **Trace/recording** | Built-in across replays, logs, and prompts — unified UI | Live viewer + session persistence for replaying auth flows | Playwright Trace Viewer (local; excellent but manual) |
| **Data retention** | 7 days (Dev) / 30 days (Startup+) | 24 hrs (Hobby) → 14 days (Pro) | Browserless: 1-30 days by plan; self-hosted: unlimited, you store |
| **Human-in-the-loop** | Not explicitly featured | ✅ Explicit HITL — human can watch/take over live session | ❌ Not built-in |
| **MTTR impact** | Best: unified debug UI reduces context-switching | Good: live takeover is unique and valuable for auth debugging | Worst for cloud; Playwright Trace Viewer is excellent locally |

---

## 4. API / Docs / Ecosystem Maturity

| Dimension | Browserbase | Steel | Self-hosted Playwright |
|---|---|---|---|
| **SDK languages** | TS, Python, Java, .NET, Go, Ruby, PHP, Kotlin, Rust (via Stagehand) | Node, Python SDKs; CDP WebSocket for any CDP tool | Native Playwright: TS, Python, Java, .NET — gold standard |
| **Framework support** | Playwright, Puppeteer, Selenium, + Stagehand (AI-native) | Playwright, Puppeteer, Selenium via CDP | Playwright only (or Puppeteer/Selenium if you configure) |
| **AI agent integrations** | LangChain, CrewAI, Mastra, MCP server (3.3k ⭐) | MCP server (47 ⭐); general CDP compatibility | None built-in |
| **Open source** | Stagehand: **22k ⭐**, 1.5k forks — massive community | steel-browser: **6.8k ⭐**, 925 forks — solid, self-hostable | Playwright itself: 70k+ ⭐ |
| **Docs quality** | Comprehensive: quickstarts, API ref, guides, operational docs, limits documented | Adequate: guides per framework, pricing/limits page, cookbook | Playwright docs are industry-leading; Browserless docs are good |
| **Compliance** | SOC2, HIPAA (BAA), DPA, SSO (Enterprise) | SOC2 mentioned on Enterprise | Browserless: SOC2, GDPR, HIPAA; self-hosted: your responsibility |
| **Maturity signal** | Strongest ecosystem, most integrations, largest community | Earlier stage but viable; open-source server is a real differentiator | Playwright is battle-tested; Browserless is mature SaaS |

> **Sources:** GitHub API (accessed 2026-04-14), docs.browserbase.com, docs.steel.dev, playwright.dev/docs, browserless.io.

---

## 5. Recommendation

### Primary: Browserbase (Startup plan, $99/mo)

**Why:**
- Best debug/observability tooling (unified replay + logs + prompt traces) → directly addresses MTTR pain
- Strongest anti-bot posture (Cloudflare partnership, dedicated stealth team, auto captcha)
- Largest SDK ecosystem aligns with TypeScript/Playwright preference
- 100 concurrent browsers handles burst patterns
- 30-day data retention for post-mortem investigation

**Risks:**
1. **Cost at scale** — overage at $0.10/hr adds up; negotiate custom pricing if usage exceeds 1,500 hrs/mo
2. **Vendor lock-in** — Stagehand API is proprietary; mitigate by keeping core logic in standard Playwright
3. **Stealth claims unverified** — no public benchmarks; must validate against your specific bot-protected targets

### Fallback: Steel (Developer plan, $99/mo)

**Why to keep it evaluated:**
- 20-40% cheaper per browser-hour at scale ($0.05-0.08 vs $0.10)
- Open-source server means escape hatch to self-host if needed
- Human-in-the-loop live viewer is uniquely valuable for auth-heavy flows
- Predictable credit model

**Why it's not primary today:**
- Shorter data retention (7 days on Developer vs 30 on BB Startup)
- Smaller ecosystem, fewer integrations, fewer SDK languages
- Lower concurrent session cap (20 vs 100) — problematic for burst patterns
- Anti-bot evidence is thinner

### Not recommended now: Self-hosted Playwright / Browserless

- Browserless is a solid product, but at 35k sessions with bot-protected targets, you'd own all stealth engineering
- Playwright Trace Viewer is excellent for local dev but doesn't provide cloud-native observability
- Operational overhead contradicts the stated goal of reducing debug turnaround time

---

## 30-Day Validation Plan

| Week | Action |
|---|---|
| **1** | Provision Browserbase Startup + Steel Developer free-tier. Run 500 sessions each against top-5 target sites. Measure: success rate, latency, captcha pass rate |
| **2** | Focus on Browserbase. Reproduce 10 recent customer failure scenarios. Measure: time-to-root-cause using session replay vs current approach |
| **3** | Load test burst pattern (peak concurrency). Confirm Browserbase handles 80+ concurrent sessions. Test Steel's 20-session cap under load |
| **4** | Cost reconciliation: project 3-month cost at actual usage. Decision gate: confirm or switch to Steel/negotiate BB custom pricing |

---

## Source Log

| Source | URL | Accessed |
|---|---|---|
| Browserbase Pricing | https://www.browserbase.com/pricing | 2026-04-14 |
| Steel Pricing/Limits | https://docs.steel.dev/overview/pricinglimits | 2026-04-14 |
| Steel Homepage (FAQ schema) | https://steel.dev/ | 2026-04-14 |
| Browserless Pricing | https://www.browserless.io/pricing | 2026-04-14 |
| Browserbase Docs | https://docs.browserbase.com/ | 2026-04-14 |
| Steel Docs | https://docs.steel.dev/ | 2026-04-14 |
| Playwright Docs | https://playwright.dev/docs/intro | reference |
| GitHub API (repo stats) | api.github.com | 2026-04-14 |

*Pricing is subject to change. All vendor anti-bot claims are unverified marketing unless otherwise noted. Validate with your own PoC.*