Direct answer: AI browser agent cost is not "price per hour" — it is browser hours + model tokens + proxy traffic + CAPTCHA added together. At Browserbase and Steel list prices, browser hours run USD 0.08–0.12; a 2-minute page task costs USD 0.0027–0.004 in browser time, and the model adds about 0.003 with a light model or 0.06 with a heavy one (20k input + 2k output). 1,000 tasks total about USD 5.67–7.00 with a light model and no proxy; add 5 MB/task of proxy traffic and that becomes another USD 30–60 — proxy and model usually cost an order of magnitude more than browser time.
This guide is for technology, operations and procurement teams using agents for back-office entry, portal lookups, reconciliation and form submission. Data cut-off: 1 November 2026; prices come from official Browserbase, Steel and OpenAI pages, in USD, excluding tax, with no FX conversion. The worked examples are editorial models, not quotes; task duration, token usage and proxy volume must be replaced with your own data.
Price per task, not per monthly plan; use fetch/search for static pages and open a browser only for interaction, login or heavy JS; proxy traffic and model choice usually matter more than browser hours.
What are the four units?
Browser infrastructure bills browser hours (Browserbase USD 0.10–0.12, Steel 0.08–0.10), models bill tokens, proxies bill GB, CAPTCHA bills per solve, and fetch/search APIs bill per 1,000 calls.
| Item | Browserbase | Steel | Note |
|---|---|---|---|
| Browser hours | Developer USD 0.12; Startup 0.10 | Launch USD 0.10; Scale 0.08 | Overage rate after included allowance |
| Proxy bandwidth | USD 12/GB (Developer); 10 (Startup) | USD 10/GB (Launch); 6 (Scale) | Heavy pages, images and video add up fast |
| CAPTCHA | Included on paid plans | USD 3/1,000 (Launch); 1 (Scale) | Appears with logins and anti-bot |
| Fetch / search | Fetch USD 1/1,000; Search 5/1,000 | Scrape USD 5/1,000 | Prefer this for static pages |
Sources: Browserbase pricing and Steel pricing (1 November 2026). Public list prices, excluding tax and plan fees; enterprise agreements are excluded.
What does one task cost?
Using Browserbase's own example ("100 browser hours ≈ 3,000 page-level tasks"), a task is about 2 minutes = 0.033 hours; browser cost is USD 0.0027–0.004, plus about 0.003 in light-model tokens, or roughly USD 0.006–0.007 per task.
| Cost line | Assumption | Browserbase Developer | Steel Scale |
|---|---|---|---|
| Browser time | 2 minutes/task | 0.033 h × 0.12 = 0.0040 | 0.033 h × 0.08 = 0.0027 |
| Model tokens (light) | 20k input + 2k output at 0.10/0.50 per 1M | 0.003 | 0.003 |
| Model tokens (heavy) | Same tokens at 2/10 per 1M | 0.060 | 0.060 |
| Proxy | 5 MB/task | 0.060 (USD 12/GB) | 0.030 (USD 6/GB) |
| CAPTCHA | 20% of tasks, one solve each | Included | 0.0006 |
Editorial model: browser time uses Browserbase's published "100 hours ≈ 3,000 page-level tasks"; model prices use OpenAI gpt-6-luna (0.10/0.50) and gpt-6.1-sol (2/10) per 1M tokens as light/heavy examples; token counts, 5 MB/task and 20% CAPTCHA are example assumptions, not benchmarks.
What do 1,000 tasks total?
With a light model and no proxy: USD 5.67–7.00. With the heavy model: 62.67–64.00. Adding 5 GB of proxy adds 30–60; CAPTCHA about 0.60. The spread is almost entirely model and proxy.
| Option | Browser | + light model | + heavy model | + 5 GB proxy | + 20% CAPTCHA |
|---|---|---|---|---|---|
| Browserbase Developer | 4.00 | 7.00 | 64.00 | 60.00 | 0 |
| Browserbase Startup | 3.33 | 6.33 | 63.33 | 50.00 | 0 |
| Steel Launch | 3.33 | 6.33 | 63.33 | 50.00 | 0.60 |
| Steel Scale | 2.67 | 5.67 | 62.67 | 30.00 | 0.20 |
Editorial model: 1,000 tasks × the per-task costs above; proxy priced on 5 GB total (5 MB/task); CAPTCHA on 200 solves (20%). Columns can be added, but plan allowances are not netted; excludes monthly fees, storage and tax.
When should you not open a browser?
For static pages, public information and read-only lookups, fetch/search is 10–100× cheaper: Browserbase Fetch USD 1/1,000, Search 5/1,000; Steel Scrape 5/1,000 — while a browser task also pays model tokens.
| Task type | Start with | Cost per 1,000 (example) | Why |
|---|---|---|---|
| Public article / product page text | Fetch / Scrape | 1–5 | No interaction; fetch HTML or content directly |
| Search-result collation | Search API | 5 | Cheaper than driving a browser and parsing |
| Login / multi-step forms | Browser agent | 2.67–4.00 + tokens | Sessions, clicks and uploads cannot be fetched statically |
| Heavy anti-bot / CAPTCHA | Browser + proxy/CAPTCHA | + 30–60 proxy | Browser hours are not the main cost |
| High-frequency read-only monitoring | Fetch first, browser as fallback | From 1 | Monitoring with a browser is the most expensive |
Editorial guidance: decide whether interaction is required, then pick the method; a single workflow can mix static fetch steps with browser steps.
Is self-hosting or open source cheaper?
Steel's browser is open source and can be self-hosted. You save the browser-hour fee but add servers, proxies, CAPTCHA and operations; 5 GB of proxy still costs 30–60, which is usually more than the browser fee.
| Cost line | How to meter | Example parameter (assumption) |
|---|---|---|
| Compute / containers | vCPU-hours or monthly cloud host | Per cloud bill |
| Concurrency and scaling | Peak concurrency × per-instance cost | 100 concurrent sessions need capacity planning |
| Proxy traffic | GB × price | 5 GB × 6–10 = 30–50 |
| CAPTCHA | Solves × price | 200 × 1–3 per 1,000 |
| Operations and upgrades | Person-days/month | 1–2 person-days |
| Compliance and isolation | One-off + quarterly | Per enterprise requirements |
Formula: self-hosted monthly cost = compute + (proxy GB × price) + CAPTCHA + (person-days × day rate) + compliance. Example parameters show magnitude, not a quote; licence and support terms follow Steel's repository.
What should procurement and acceptance check?
Cost per task, session length limits, concurrency, whether proxy and CAPTCHA are included, data retention, and failure/retry billing — not just the browser-hour rate or plan fee.
| Acceptance item | Suggested definition | Evidence to keep |
|---|---|---|
| Cost per task | (Browser + model + proxy + CAPTCHA) ÷ successful tasks | Invoice, task logs |
| Session limit | Max session length and timeout billing | Plan details, load test |
| Concurrency | Peak concurrency and queueing | Load-test record |
| Proxy and CAPTCHA | Included or overage price | Price list, invoice lines |
| Data retention | Recordings, logs and screenshots retention | DPA, retention policy |
| Failure billing | Timeouts, crashes, CAPTCHA failures | Contract, sample invoice |
| Portability | Playwright/Puppeteer basis, ability to switch vendor | Code, API docs |
| Compliance | Credentials, personal data, target-site terms | Review records, authorisation |
Editorial guidance, not legal advice; operating third-party portals on a customer's behalf requires checking the target site's terms and the scope of authorisation.
Who should read this — and who should not?
For teams doing back-office entry, portal lookups, reconciliation and forms. Not for purely static scraping, or projects that need to bypass someone else's account permissions.
Read if: you are in engineering, operations or procurement; you are comparing Browserbase, Steel or self-hosting; you need a budget for 1,000 tasks. Skip if: you only scrape public static pages; you lack authorisation for the target site; you expect an agent to bypass verification and permissions. Supports: per-task cost modelling, browser versus fetch choices, proxy budgeting and acceptance criteria.
Lingyi does not rank browser infrastructure or promise task success. To connect browser agents to internal systems, permissions and audit, see LingHarness; to fold agent cost into project budgets, talk to us.
Related reading: AI data collection cost, AI agent implementation guide, and agent operations metrics; this article only covers browser-agent hourly/per-task cost and selection.
FAQ
How is an AI browser agent billed?
Browser hours (Browserbase USD 0.10–0.12, Steel 0.08–0.10), model tokens, proxy GB and CAPTCHA solves are billed separately; fetch/search APIs bill per 1,000 calls. Add all four before computing cost per task.
What does one task cost?
At 2 minutes per task with a light model (20k input + 2k output): browser USD 0.0027–0.004 plus about 0.003 in tokens, roughly 0.006–0.007 per task; a heavy model pushes it to about 0.06.
Why is the proxy more expensive than the browser?
Because proxies bill per GB: 5 MB/task × 1,000 tasks = 5 GB, which at USD 6–12/GB is 30–60, while browser time is only 2.67–4.00. Images, video and heavy pages scale it further.
Do static pages need a browser?
No. Use Fetch/Scrape for public content (USD 1–5 per 1,000) and Search APIs for search (USD 5 per 1,000); open a browser only for login, clicks, uploads or heavy JS rendering.
Is self-hosting cheaper?
Not necessarily. Steel is open source and self-hostable, saving the browser-hour fee, but proxies, CAPTCHA, concurrency capacity and operations remain; 5 GB of proxy alone costs 30–50, usually more than the browser fee.
Are failures and retries billed?
Rules vary: browser hours are usually metered on actual runtime, so timeouts and crashes still consume time, and CAPTCHA is billed per solve. Ask about timeouts, crashes, CAPTCHA failures and retries before signing.
How should we estimate our own bill?
Run 100–200 real tasks, record average duration, tokens, proxy volume, CAPTCHA solves and success rate, then apply the rates; do not rely on browser hours × rate alone.
Method and limitations
This article uses only official pages accessible on 1 November 2026: Browserbase pricing, Steel pricing and OpenAI API pricing. All amounts are converted from public list prices and published examples; no enterprise agreements or reseller quotes are used. USD is not converted to RMB and no vendor ranking is implied. Plan allowances, regional prices, CAPTCHA and proxy rules can change; Browserbase's "100 hours ≈ 3,000 page-level tasks" is its own example, used here to derive 2 minutes per task. Token counts, 5 MB/task, 20% CAPTCHA and 1,000 tasks are example assumptions, not benchmarks; the model prices are two OpenAI tiers used to illustrate light and heavy usage, not the only options. The article does not assess task success, stability or compliance and is not legal advice.
Sources and basis
Data cut-off: 1 November 2026. The following are official pages; prices, allowances and terms follow the original text and invoice.
- Browserbase, Pricing: browserbase.com/pricing. Browser hours USD 0.12/0.10 (Developer/Startup), proxies 12/10 per GB, Fetch USD 1/1,000, Search 5/1,000; "100 browser hours ≈ 3,000 page-level tasks".
- Steel, Pricing: steel.dev/pricing. Browser hours USD 0.10/0.08 (Launch/Scale), proxies 10/6 per GB, CAPTCHA 3/1 per 1,000 solves, Scrape USD 5/1,000; open-source and self-hosting notes.
- OpenAI, API pricing: platform.openai.com/docs/pricing. gpt-6-luna at 0.10/0.50 and gpt-6.1-sol at 2/10 per 1M tokens, used as light/heavy examples.