This site contains affiliate links.
If you sign up through one of those links I may earn a commission, at no extra cost to you. I’ve kept the recommendations honest — a tool that pays well but fits badly is worth nothing to either of us.
Six pricing pages, six different units, none of them comparable
Put these tools side by side and you’ll notice something before you notice anything else: no two of them sell the same thing.
Browse AI sells credits. Octoparse sells tasks — which are saved configurations, not runs. Apify sells compute units, meaning one gigabyte of RAM for one hour. Bright Data sells records. Thunderbit sells credits too, but a Thunderbit credit is one row where a Browse AI credit is ten. Firecrawl sells credits that mean pages, except when you turn on structured extraction, at which point one page costs five.
You cannot rank these by price. There is no shared denominator. And that’s not an accident of a young market — the unit is the product decision. It determines who the tool is for, what it’s cheap at, and where your bill detonates.
So this comparison is organised around the unit rather than the sticker. I’ve converted everything to the same question wherever the vendor documents enough to allow it: what does it cost to extract a thousand rows from a site that doesn’t want you to? Where a vendor makes that impossible to compute, I say so, because that’s a finding about the product and not a gap in the research.
Three questions that decide this
1. Are you fetching, or are you watching?
These are different jobs and the tools split cleanly along them.
Fetching is one-off or occasional bulk extraction. You want ten thousand rows out of a site, once, or once a quarter. What matters is throughput and cost per row.
Watching is monitoring a known set of pages for change — a competitor’s prices, a job board, a set of listings — and being told when something moves. What matters is scheduling, alerting, and the cost of running the same small job over and over.
The trap is that watching is priced as repeated fetching. Browse AI documents this plainly: monitoring means each check is a separate task, so watching ten pages hourly is ten credits an hour, not ten credits. A job that looks tiny — “just tell me when this price changes” — is the most expensive shape of work in this category, because you’re paying for the same extraction hundreds of times.
If you’re watching, price the cadence, not the page count. Ten pages checked every fifteen minutes is 28,800 checks a month.
2. What does one unit actually buy — and is that documented at all?
Here is the honest ranking, best to worst, and it’s the single most useful thing in this article:
Browse AI publishes the conversion. Ten rows of data equal one credit. One screenshot is one credit. There’s a floor of one credit per task, and a set of named “premium sites” — Amazon, Walmart, Instagram, LinkedIn, X, Google among them — that cost two to ten credits per task instead. That last part matters more than it looks, and I’ll come back to it.
Bright Data sells records, billed on successful delivery. One record is one row. That’s the cleanest unit here, and it’s the only one where a non-technical buyer can multiply two numbers and get a bill.
Firecrawl documents its credits in detail — one credit per page scraped or crawled, two per ten search results, plus four per page if you turn on structured JSON extraction. But the credit cost of its enhanced stealth proxy mode is nowhere in the documentation, and the automatic mode silently retries with it on failure. You cannot compute the cost of scraping a protected site with Firecrawl from published information.
Thunderbit documents one row per credit — clean, until you find data enrichment at thirty credits per successful query. Enriching five hundred rows costs fifteen thousand credits, which is five months of the Pro tier’s monthly allowance.
Octoparse doesn’t meter volume at all above the free tier. A task is a saved crawler, not a run and not a quantity of data. Paid plans say “unlimited data export.” Your real constraint is concurrency and wall-clock time. This is genuinely different and it’s the reason to pick it.
Apify cannot document the conversion, and never will be able to. A compute unit is a gigabyte of RAM for an hour. How many rows that produces depends entirely on which Actor you run and how heavy the target page is — and each Actor in the store sets its own price on top, per-event or per-result or as a monthly rental. This isn’t Apify being cagey. It’s a structural property of being a marketplace, and you should price it as a variable cost you’ll only know in arrears.
3. Who fixes it when the site changes?
Every scraper breaks. Sites get redesigned, class names change, a login wall appears. The question is what happens next.
With a recorder like Browse AI or Octoparse, you fix it — you re-record the selectors. That’s fine when you own five robots and awful when you own fifty.
With Apify, someone else maintains it, if you’re renting a popular Actor with an active author. That’s a real advantage nobody advertises, and it inverts when the Actor is abandoned.
With Bright Data’s scraper APIs, the vendor maintains it for the sites they cover, and you’re billed only on successful delivery — so a break costs you nothing directly, it just stops producing.
With Firecrawl, there’s nothing to break in the selector sense, because you’re not selecting; you’re asking for the page as markdown. Different failure mode: the site blocks you, or the extraction returns something subtly wrong.
Ask yourself how many distinct targets you’ll be scraping in a year. Under five, a recorder is fine. Over twenty, maintenance is your dominant cost and you want someone else doing it.
Before any of this: is the scraping itself okay?
Worth thirty seconds. Scraping publicly visible pages is broadly lawful in a lot of jurisdictions, but that’s not the whole question. Check the site’s terms of service and its robots.txt, don’t scrape behind a login you agreed not to automate, don’t take personal data you have no basis to hold, and rate-limit yourself so you’re not degrading someone’s site.
The practical version: if you’d be uncomfortable telling the site owner what you’re doing, that’s the signal — not a legal opinion, but a reliable heuristic. And note that Thunderbit’s whole design runs inside your own logged-in browser session, which sidesteps some technical barriers precisely because it is you. That’s convenient and it does not make the terms of service go away.
The comparison
Prices verified against vendor pricing pages on 27 August 2026. This category changes fast — verify before you commit, and screenshot whatever you rely on.
| Tool | What it is | The unit | Realistic entry plan | Free tier | The thing that gets people |
|---|---|---|---|---|---|
| Browse AI | Point-and-click recorder + change monitoring | Credits — 10 rows = 1 credit, documented | Personal $19/mo annual · Professional $69/mo annual | 50 credits/mo ≈ 500 rows | Premium sites cost 2–10 credits per task, not 1 |
| Octoparse | Desktop visual scraper + cloud runs | Tasks = saved crawlers. No volume metering on paid | Standard $99/mo annual ($119 monthly) | 10 tasks, local only | Free tier has no cloud, no scheduling, no anti-blocking at all |
| Thunderbit | AI browser extension, runs in your session | Credits — 1 row = 1 credit | Starter $9/mo annual · Pro $16.50/mo annual | 6 pages/month | Enrichment costs 30 credits per query |
| Apify | Marketplace of pre-built scrapers | Compute units (1 GB RAM × 1 hr) + per-Actor pricing | Starter $29/mo + usage | $5/mo of usage | CU-to-rows is undocumentable by design |
| Bright Data | Proxy and data infrastructure | Records, billed on success | Pay-as-you-go from $1.50/1k records | 5,000 records/mo, no card | Not a no-code recorder — there’s no builder |
| Firecrawl | Crawl API returning clean markdown | Credits — 1 page = 1 credit, +4 with JSON | Hobby $16/mo annual · Standard $83/mo annual | 1,000 credits/mo | Structured extraction quintuples your per-page cost |
Tool by tool
Browse AI
What it’s for: Watching a known, bounded set of pages on a schedule, without writing anything.
You record a robot by clicking through a page the way you would if you were doing it by hand, and it replays in their cloud on whatever schedule you set. Monitoring and alerting are first-class rather than bolted on, which is what separates it from the desktop scrapers.
It also has the best-documented unit economics in this comparison, so let me use it to show you what “read the unit” means in practice.
Ten rows equal one credit. But there’s a minimum of one credit per task, so a robot that returns a single row costs the same as one returning ten — ten times the marginal rate. And then there’s the premium-sites list: Amazon, Walmart, Costco, Home Depot, Instagram, LinkedIn, X, Google and others cost two to ten credits per task minimum. Which means that on the sites people overwhelmingly actually want to scrape, your effective capacity is somewhere between half and a tenth of the number printed on the pricing page. That’s not a gotcha — Browse AI publishes the list openly, which is more than most do. But you have to go and read it.
Now the billing thing, and I want to be precise here because I got it wrong the first time and had to correct myself.
On the Personal tier, monthly billing quotes 2,000 credits per month, and the cheapest annual plan quotes 12,000 credits per year. Same slot on the page, same typeface — different unit. Twelve thousand a year is a thousand a month, so the entry annual plan gives you half the monthly capacity for 60% less money. There is a parity option at $38/mo annual, which gets you 24,000/year. And on Professional the two ladders match exactly — 5,000/month monthly, 60,000/year annual, right up the scale.
So this is not a company-wide trick, and I’d be misleading you if I framed it as one. Annual billing is never worse per credit. What it is, is a unit change presented as a number change, confined to the entry tier, on a page most people scan rather than read. That’s worth knowing about, and it’s worth checking on every tool you evaluate, because Thunderbit does the same thing on every one of its tiers and Firecrawl deliberately doesn’t do it at all.
Pros
- Ten-rows-per-credit is published, which makes this one of the few tools here where you can forecast a bill.
- Monitoring and scheduled re-runs are the core product, not an afterthought.
- Unlimited robots on every tier, including free.
- The free tier is the full platform at 50 credits a month, not a feature-crippled demo.
- The premium-sites surcharge is disclosed openly rather than discovered on an invoice.
Cons
- The one-credit floor per task punishes small, frequent jobs — which is exactly the shape of monitoring work.
- Premium sites cut real capacity by 2–10×, and they’re the sites you probably want.
- Websites are capped per tier (2 / 5 / 10) with extras billed separately, so the credit number isn’t the only limit.
- The review base is thin: 62 on G2, 63 on Capterra, against Apify’s 596 and Bright Data’s 341. Ratings are good, but there’s much less accumulated public experience to lean on.
- Personal has no higher monthly SKU, so the only exit from 2,000 credits is a jump to Professional.
Octoparse
What it’s for: People who want volume without a meter running, and will pay for that with a desktop app and a learning curve.
Octoparse is the odd one out here and it’s the most interesting deviation in the category. Above the free tier it does not meter data volume at all. A “task” is a saved crawler configuration — not a run, not a row count. Standard gives you 100 of them, Professional 250, and the pricing page says unlimited data export. Your constraints are concurrency (6 or 20 cloud processes) and how long things take.
If your work is “extract a lot from a few sites,” that model is dramatically better than credits, and nobody else offers it.
Two warnings. First, the free tier does not predict paid behaviour at all — it’s local-execution only, with no cloud, no scheduling, and no anti-blocking whatsoever. The moment your target has a CAPTCHA or you need an unattended overnight run, you’re at $99/mo on annual billing or $119 monthly. There is nothing in between, and it’s the steepest entry cliff in this comparison.
Second, their pricing page contradicts itself. The page title still advertises “from $69/Month” while the plan cards say $119 and $99. I’ve cited the cards. Screenshot whatever you rely on, because this page moves.
Also worth knowing: proxies, CAPTCHA solving and pay-per-result templates are metered separately on top of the plan — residential proxies at $3/GB, CAPTCHA solving at $1–1.50 per thousand. And the existence of pay-per-result templates “for complex or highly protected websites” is a quiet admission that the normal builder struggles on those targets.
Pros
- No volume metering above Free — genuinely unusual and genuinely valuable at scale.
- Mature product with real depth for complex, multi-step extraction.
- 52 G2 and 106 Capterra reviews, so there’s accumulated public experience.
- Anti-blocking and CAPTCHA solving are real, on paid tiers.
Cons
- The free-to-paid cliff is $0 to $119/mo with nothing between, and the free tier can’t do the things you’re evaluating.
- Desktop app, Windows/Mac install, steepest learning curve here.
- Add-ons are metered separately, so “unlimited” is unlimited rows, not unlimited spend.
- The pricing page disagrees with its own title tag, which is not confidence-inspiring.
Thunderbit
What it’s for: Small ad-hoc extraction from sites you’re already logged into.
A Chrome extension where you describe what you want in plain language and it runs inside your own browser session. That last part is the whole design: it isn’t fighting a login wall or a fingerprinting check, because it genuinely is you, in your browser, with your cookies.
For gated content that’s a real advantage no proxy-based tool can match. For anything adversarial at volume, it’s the weakest option here — no proxy rotation or CAPTCHA solving is published on any tier.
One row is one credit, documented. Sub-page scraping doubles it. And enrichment costs thirty credits per successful query, which is the number that ruins budgets: enriching a five-hundred-row list costs fifteen thousand credits, and Pro’s monthly allowance is three thousand.
Thunderbit also runs the same monthly-versus-annual unit switch as Browse AI, but more consistently: Starter is 500 credits/month on monthly billing versus 5,000 per year on annual — a 17% reduction in effective allowance. Pro is 3,000/month versus 30,000/year, the same 17%. Annual is still cheaper per credit. It’s still a smaller allowance than the monthly number implies.
Pros
- Running in your own session solves login-gated extraction elegantly.
- Cheapest paid entry here at $9/mo on annual billing.
- One row per credit is documented and easy to reason about.
- Free exports to Sheets, Airtable and Notion.
Cons
- Six pages a month on free is a demo, not a tier.
- Enrichment at 30 credits a query is an order-of-magnitude cost most people won’t see coming.
- No published anti-bot capability of any kind.
- 11 G2 and 13 Capterra reviews — at that sample size the ratings mean very little.
Apify
What it’s for: Getting one specific popular target scraped by someone else’s well-maintained code.
Apify isn’t a scraper, it’s a platform and a marketplace. Thousands of pre-built Actors cover the common targets, and if the site you want is popular, someone has already solved it and keeps solving it as the site changes. That maintenance transfer is the real product.
The cost model is the price of that. A compute unit is a gigabyte of RAM for an hour, billed at $0.20 falling to $0.13 at higher tiers. Residential proxies are $8/GB, datacenter $1/IP, data transfer $0.20/GB, and store Actors set their own pricing on top. Your monthly plan fee is prepaid usage credit, not an allowance, and it doesn’t roll over.
I want to be fair about this rather than treat it as a flaw: the conversion from compute units to rows is not undocumented, it is undocumentable. It depends on the Actor and the target. No honest vendor could publish that number. But the consequence for you is real — this is the one tool here where you genuinely cannot know your bill in advance, and if predictability matters more than capability, that’s disqualifying.
Pros
- By far the largest review base here — 596 on G2, 538 on Capterra — so you can research an Actor before committing.
- Someone else maintains the scraper for popular targets.
- Enormous breadth; if it’s a well-known site, it’s covered.
- Real infrastructure underneath, with proper concurrency and scaling.
Cons
- You cannot forecast cost, structurally.
- Prepaid usage is use-it-or-lose-it monthly.
- Actor quality varies with the author, and an abandoned Actor is worse than no Actor.
- Proxies, transfer and storage are all metered on top of compute.
Bright Data
What it’s for: Volume, hostile targets, and knowing your cost per row in advance.
A caveat first, because including it in a no-code roundup needs one: Bright Data is infrastructure, not a recorder. There’s no point-and-click builder. If you wanted to click a page and get a spreadsheet, this isn’t it.
What it is instead is the reference implementation for the hard part. Web Unlocker names Akamai Bot Manager, hCaptcha and reCAPTCHA v3 among what it handles, and bills only on successful delivery. Scraper APIs return records — one record, one row — from $1.50 per thousand on the product page. And the free tier is 5,000 records a month with no card, which is by a wide margin the most generous here.
Two things to watch. The two pricing pages disagree: the main index says $0.75 per thousand records, the Web Scraper API product page says $1.50. I’ve used the product page figure and you should verify which applies to you. And the $499/mo Scale commitment only beats pay-as-you-go above roughly 333,000 records a month — below that, PAYG is cheaper.
Pros
- Records billed on success is the cleanest, most forecastable unit in this comparison.
- Genuinely handles anti-bot, at scale, as its core competence.
- 5,000 free records a month, no card.
- 341 G2 reviews and a long operating history.
Cons
- No visual builder — this is a caveat for a no-code reader, not a defect.
- Its own two pricing pages quote different per-record rates.
- The Scale plan is a worse deal than PAYG below ~333k records/month.
- The proxy side raises real diligence questions about how the network is sourced; worth reading up on before you commit at volume.
Firecrawl
What it’s for: Feeding clean page content into an LLM pipeline — and it’s here as the contrast case, because it isn’t no-code at all.
You call an API and get markdown back. No builder, no recorder, no clicking. If you’re not writing code, stop reading this section.
I include it for two reasons. First, because “I want website data” increasingly means “I want website data in a RAG index,” and this is what that looks like. Second, because Firecrawl is the control case that proves the billing pattern elsewhere is a choice. Its credit allowances are identical on monthly and annual billing — only the price changes. Browse AI and Thunderbit could have done that and didn’t.
The thing that catches people is the modifier, not the tier. One credit per page, but turn on JSON extraction and it’s five — so the Hobby plan’s 5,000 credits become 1,000 pages the moment you use the feature most people bought it for. Prompt-injection checking adds four more, and five credits are charged even on a failed scrape. And the credit cost of enhanced stealth proxy mode is not documented anywhere, while auto mode retries into it on failure.
Pros
- The most granular published credit documentation in this comparison.
- Credit allowances don’t change between monthly and annual billing.
- 1,000 free credits a month, no card.
- Purpose-built for LLM ingestion, and good at it.
Cons
- Not no-code. At all.
- JSON extraction quintuples per-page cost, and that’s the normal use case.
- Enhanced proxy cost is unpublished, so protected-site workloads can’t be budgeted.
- One G2 review and no Capterra profile — there’s essentially no public track record yet.
A note on ParseHub
ParseHub still appears in most roundups and it’s still running — pricing page live, copyright current, no shutdown notice. But Standard is $189/mo and Professional is $599/mo, prices that haven’t moved in years, with no published CAPTCHA handling and the weakest anti-bot story here. Free projects are also public: “public project designs and the data from their runs may be shared with other users,” which rules it out for anything competitive or client-facing.
I’m not going to tell you it’s dead, because I found no evidence for that. I’d just say the highest entry price in the category should come with the best capabilities, and here it comes with the weakest.
Which one, for which situation
You’re watching a handful of pages for changes and you don’t write code. Browse AI. Monitoring is the product rather than a feature, and the credit conversion is published so you can forecast. Check the premium-sites list before you commit — if your targets are on it, halve your capacity estimate at best.
You need a lot of rows out of a few sites and you hate meters. Octoparse, for unmetered volume above the free tier. Budget for $99/mo on annual billing, not the $69 the page title still claims, and don’t evaluate on the free tier — it can’t do cloud, scheduling or anti-blocking.
Your target is a well-known site and you’d rather not maintain a scraper. Apify. Rent someone’s Actor, check its review history first, and accept that you’ll learn your bill after the fact rather than before.
The site is fighting you, or you need cost per row to be a known number. Bright Data. Stay on pay-as-you-go until you’re past roughly 333,000 records a month — the Scale commitment is worse below that.
You need data from behind a login you legitimately have. Thunderbit, because running in your own session is the cleanest answer to that specific problem. Avoid it for enrichment work, where 30 credits a query will eat a month’s allowance in an afternoon.
You’re building a RAG pipeline and you write code. Firecrawl. Price it with the JSON modifier on, because that’s how you’ll actually use it.
You want to learn whether this is worth paying for at all. Bright Data’s 5,000 free records a month is the most generous no-card free tier here, and Browse AI’s 50 credits is the least crippled free version of a full product. Between them you can answer the question for nothing.
The habit worth building
The specific findings in this article will age. Prices move, tiers get renamed, Octoparse’s page title will presumably get fixed eventually.
What won’t age is the habit: on any pricing page in this category, find the unit before you look at the number. Then ask three follow-ups. Is that unit per month or per year — and is it the same unit in both billing modes? Is there a floor or a minimum charge that makes small jobs cost more than they should? And is there a surcharge class — premium sites, structured extraction, enrichment, stealth proxies — that applies to the exact thing you’re planning to do?
Every one of the six tools here has at least one of those three. Browse AI has a floor and a surcharge class. Thunderbit has a surcharge class that costs thirty times the base rate. Firecrawl has one that quintuples it. Octoparse has no unit at all above free, which is its own kind of answer.
None of that is hidden. It’s all published, in help docs, one click off the pricing page. It’s just not on the page with the big numbers on it — and that’s the part worth remembering next time.