Firecrawl Review (July 2026): Three New Tools, One Rating Revisited
Published July 1, 2026 · Updated July 14, 2026 · by Pondero Reviews
The short version
Firecrawl added Lockdown Mode, /parse, and question and highlights on top of /monitor. Pricing rechecked 2026-07-14, unchanged. Monitor frequency, not page count, is what picks your tier.
Pros
- ✓Markdown-by-default output still lands in the shape an LLM context window wants, and v2.11's deterministicJson returns structured JSON without an LLM call for cheaper, repeatable extraction per the v2.11.0 release notes
- ✓Keyless access means an agent connects to the Firecrawl MCP and starts scraping with no key-generation step, using the 1,000 free monthly credits per the keyless launch post
- ✓Pricing held flat since May: Hobby $16/mo, Standard $83/mo billed yearly, per firecrawl.dev/pricing fetched 2026-07-01
- ✓/monitor delivers only page diffs on a schedule, cutting up to 90% of LLM tokens versus full re-ingestion per the changelog, at 1 credit per page per check
- ✓Research Index leads arXivQA recall at 53.3% versus 45.4% next-best at similar cost per the launch benchmark, a real edge for AI/ML research agents
- ✓The question and highlights formats return a grounded answer or verbatim excerpts from a page using up to 100x fewer tokens than a full scrape per Firecrawl's launch post, so the saving lands on your model bill
Cons
- ✕Credit math still surprises crawl-heavy workloads: a 5,000-page crawl is 5,000 credits, which clears the entire Hobby tier in one run
- ✕The Research Index is niche. It indexes arXiv plus research GitHub, so general web-scraping and non-research knowledge bases get nothing from it
- ✕Agent endpoint pricing is dynamic after 5 free daily runs per the pricing page, so a heavy agentic workload has no fixed line-item you can budget in advance
- ✕PII redaction via redactPII strips names, emails, and secrets per the release notes, but whether that clears your GDPR obligations is a legal-team call, not a checkbox the docs answer
- ✕Concurrency stays low on the cheap tiers: 5 concurrent on Hobby, so a real load test still needs Standard's 50 concurrent
- ✕Firecrawl publishes no credit cost for Lockdown Mode, question, highlights, or the new web-scale /monitor, so four of the newer endpoints can't be budgeted before you run them
Firecrawl Review (July 2026): Three New Tools, One Rating Revisited
You found Firecrawl, the pricing page loads, the docs read clean, and you are staring at two buttons: Hobby at $16 a month or Standard at $83, both billed yearly per firecrawl.dev/pricing (fetched 2026-07-01). That is the real decision, and the answer hasn't changed much since our May review. Start on the free tier or Hobby to prove the pipeline, then jump to Standard the moment your monthly credit burn clears about 6,000. What has changed is that the free step now needs no API key at all, and the ceiling on what you can do before upgrading is higher.
Three things shipped in the 43 days since that review. Firecrawl added /monitor for change-tracking, a Research Index for AI/ML literature search, and v2.11.0, which made the core endpoints keyless and added automatic PII redaction. None of them move the tier you buy. They change how far the tier you buy takes you. That is the distinction this piece is about.
We compared the current pricing against real agent workloads, checked the new features against the docs, and looked at where Firecrawl sits versus Apify and ScrapingBee at scale. The short version: the product got broader without getting more expensive, so the rating nudges up. Rated 4.5 of 5, from 4.4 in May.
What the May review said, and what actually changed
The May verdict was a 4.4 and a Hobby-tier recommendation: buy Hobby for a real prototype, move to Standard past roughly 10,000 pages a month, and budget for the credit-math conversation on any crawl-heavy day. The pricing deep-dive from June walks the tier-by-tier math if you want it. The core thesis held: Firecrawl optimizes for the shape of output an LLM ingests, where ScrapingBee and Apify optimize for raw HTML and configurable actors.
Then three changes landed. What each one actually does for you:
/monitor (May 26). You give it a URL and describe what to track in plain English ("alert me when the Claude Code docs add new slash commands"), and it fires a signed webhook or an email when the page changes. It delivers only the diff, which the changelog says cuts up to 90% of LLM tokens versus re-ingesting the whole page every check (firecrawl.dev/changelog, fetched 2026-07-01). Cost is 1 credit per page per check. For an agent that watches 50 competitor pages daily, that is a predictable 1,500 credits a month, and your model never re-reads text that didn't move.
Web-scale /monitor then landed on July 1, the same day this review first published. Firecrawl's post says /monitor "now watches the entire web, pinging you or your agent the moment something relevant comes online," which moves it from watching a page or a site you name to standing over the open web (firecrawl.dev/blog/web-scale-monitor). Cadence options and the credit cost of that web-scale tier are not published, so the 1-credit-per-page-per-check math below covers the page-level version and nothing more.
Research Index (June 16). A specialized index over 3M+ arXiv papers plus GitHub artifacts (issues, merged PRs, READMEs) from top research repos, refreshed daily. On arXivQA it hits 53.3% recall at $0.32 per task against 45.4% for the next-best provider, and scores 0.750 MRR, meaning the right paper usually lands in the top two results (firecrawl.dev/blog/research-index-launch, fetched 2026-07-01). It runs through a /search/research endpoint plus the CLI, MCP, and SDKs. We covered how to wire it into an agent in the Research Index agent guide.
v2.11.0 (late June). The core endpoints (/scrape, /search, /interact, /parse) now work with no API key when called from the official MCP, CLI, or SDK clients, drawing on 1,000 free credits a month per the keyless launch post. The same release added redactPII (strips names, emails, phone numbers, addresses, and secrets before content returns), deterministicJson (structured JSON from cached extractors, no LLM call per request), and video discovery on any page (github.com/firecrawl/firecrawl/releases, fetched 2026-07-01).
The operational read: keyless kills the signup-before-you-build friction, which matters most for coding agents that would otherwise stall waiting for a human to paste a key. Wiring Firecrawl into Claude Code is now a single line with no key step, per the keyless launch post:
# Add Firecrawl MCP to Claude Code, no API key required
claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcpdeterministicJson is the sleeper. Structured extraction that skips the LLM is both cheaper and repeatable, so the same page returns the same JSON shape every run. The Research Index is a set of /search/research endpoints, and the paper search runs keyless out of the box, per the docs:
# Keyless; add -H "Authorization: Bearer $FIRECRAWL_API_KEY" for higher rate limits
curl -s "https://api.firecrawl.dev/v2/search/research/papers?query=diffusion%20image%20synthesis&k=20"The response returns ranked papers with a canonical paperId, title, abstract, and score, and sibling endpoints cover related-paper lookup and GitHub history (docs.firecrawl.dev/features/research, fetched 2026-07-01). Run the same call through the CLI, MCP, or an SDK and it slots into whatever harness you already have.
Four capabilities the last review skipped
Four more things shipped between April 28 and May 26, bracketing the window above rather than sitting inside it, which is how both reviews walked straight past them. /parse got a passing mention in the keyless endpoint list and nothing else. Each one moves who Firecrawl is for, so here is what each does, what it costs, and when to bother.
Lockdown Mode (April 30). Scrapes resolve against Firecrawl's cache with, in the launch post's words, "no outbound request, zero data retention" (firecrawl.dev/blog/lockdown-mode). One flag turns it on across every SDK, the CLI, and MCP. What a lockdown scrape costs in credits isn't published, so carry it as an unknown when you model spend. Skip it unless an outbound request to a third-party site is itself a compliance event, which describes a lot of finance and healthcare engineering.
/parse (April 28). Hand it a local file and it runs through the same parsing engine behind /scrape. It accepts PDF, DOCX, DOC, ODT, RTF, XLSX, XLS, and HTML up to 50 MB, and hands back markdown with structure, reading order, and tables preserved, JSON against a schema you supply, or a summary (firecrawl.dev/blog/introducing-parse). Cost is quoted as the "same credit model as /scrape: one call plus any LLM formats you request," with no fixed per-page figure published. Worth it if you are dropping PDF whitepapers into a RAG pipeline and would rather not own a PyMuPDF wrapper for the rest of the project's life.
question and highlights (May 8). Ask question for a grounded answer from a page and you get the answer; call highlights and you get verbatim excerpts instead of the page. Both run on a managed LLM stack with prompt-injection hardening built in, and Firecrawl's post claims they use "up to 100x fewer tokens than a full scrape" (firecrawl.dev/blog/introducing-question-and-highlights). Read that 100x as a vendor-reported ceiling. A short page that is mostly the answer you wanted saves you almost nothing, and the credit cost of these formats is again undocumented. Worth it if model tokens, not Firecrawl credits, are the line item that actually hurts, which for most RAG pipelines they are.
Vercel Marketplace (May 26). One click installs Firecrawl into a Vercel project and the API key gets injected into the project environment for you (firecrawl.dev/blog). The catch is downstream: billing flows through Vercel, which is fine right up until your org routes vendor invoices separately and finance wants to know why scraping is buried in the hosting bill. Only pays off if you already deploy on Vercel and skipping one pasted secret beats owning the invoice.
Pricing math, updated for July 2026
The dollar figures below are identical to May, so anyone who priced this out six weeks ago can trust their old numbers. A re-check on 2026-07-14 found every tier unchanged: same prices, same credit allowances, same concurrency caps. All rows are from firecrawl.dev/pricing, fetched 2026-07-14:
| Tier | Monthly (billed yearly) | Credits/mo | Concurrent |
|---|---|---|---|
| Free | $0 | 1,000 | 2 |
| Hobby | $16 | 5,000 | 5 |
| Standard | $83 | 100,000 | 50 |
| Growth | $333 | 500,000 | 100 |
| Scale | $599 | 1,000,000 | 150 |
Credit consumption is the number that decides your tier, not the page count. Scrape, Crawl, and Map cost 1 credit per page. Search is 2 credits per 10 results. Interact is 2 credits per browser minute. Monitor is 1 credit per page per check. The Agent endpoint gives you 5 free runs a day, then switches to dynamic pricing (firecrawl.dev/pricing, fetched 2026-07-01).
Map that to a workload. A research agent that scrapes 200 pages every morning burns 200 credits a day, which is 6,000 a month. That sits just over the 5,000-credit Hobby ceiling, so Standard is the tier, and Standard gives you 100,000 credits of headroom before you think about it again (firecrawl.dev/pricing, fetched 2026-07-01). This is the upgrade trigger: not a page count, but the moment your steady-state burn crosses roughly 6,000 credits. Below that, Hobby covers you for $16. Above it, you're on Standard whether you crossed by a little or a lot, so there's no penalty for growing into the tier.
Now the part that catches people. Monitor bills 1 credit per page per check (firecrawl.dev/pricing, fetched 2026-07-14), so the check interval multiplies straight through your page count, and the interval is the variable nobody puts in the spreadsheet. Take the same 50 competitor pages from above and watch them hourly instead of daily. 50 x 24 is 1,200 credits a day, roughly 36,000 a month, against 1,500 a month at a daily cadence: a 24x swing with no extra page added. It is also more than seven times the entire 5,000-credit Hobby allowance, so an hourly monitor burns through Hobby before the first week is out, and on Standard it eats 36,000 of your 100,000 credits, leaving about 64,000 for actual scraping. Slow the cadence and widen the watch and the math relaxes: 200 pages every six hours is 200 x 4, or 800 credits a day, about 24,000 a month, which leaves roughly 76,000 on Standard. That is all arithmetic off the published per-check rate, so run it against your own cadence before you pick a plan. Frequency picks your tier. Page count barely gets a vote.
If your workload touches user data, v2.11's PII redaction is a one-flag change on the same scrape call, per the docs:
from firecrawl import Firecrawl
firecrawl = Firecrawl(api_key="fc-YOUR-API-KEY")
# redactPII strips names, emails, phones, addresses, and secrets from the markdown
doc = firecrawl.scrape("https://example.com/contact", redact_pii=True)
print(doc.markdown)Set redactPII: true (the SDKs expose it as redact_pii) and the returned markdown comes back with PII stripped (docs.firecrawl.dev/features/scrape, fetched 2026-07-01). Whether that satisfies your GDPR or CCPA obligations is a legal-team question, not a checkbox the docs answer. It reduces what lands in your logs and your vector store; it does not certify compliance.
The Interact and Agent math is where budgets slip. A page that needs a click-and-wait before the content loads runs through Interact at 2 credits per browser minute, so a slow interactive page can cost more than a plain scrape by an order of magnitude. And the Agent endpoint's dynamic pricing after 5 daily runs means an agentic workload has no fixed line item you can forecast. If you lean on Agent or Interact heavily, model those separately from your scrape credits before you commit to a tier, because the flat credit-per-page math undersells them.
Research Index: when it earns its place, and when to skip it
The Research Index is not a general Firecrawl upgrade. It targets one job: retrieval over AI/ML research. If you are building a literature-review agent, a research copilot, or anything that pulls papers and their implementing code, the 53.3% recall on arXivQA is a measurable edge, and the daily refresh means an agent can surface a paper published this week (firecrawl.dev/blog/research-index-launch, fetched 2026-07-01). It plugs into Codex, Claude Code, and Grok Build through the same MCP an agent already talks to, so adoption is a config change, not a rebuild.
Skip it if your scraping is general web data, e-commerce, competitive intel, or an enterprise knowledge base where arXiv is irrelevant. The index adds nothing to those workloads and costs you the attention of evaluating a feature you'll never call. The honest cut: this is a bet on the AI/ML research vertical, and it is a strong one if you're in that vertical. For everyone else, it's a headline that doesn't apply. Our web-search API comparison covers the general Search endpoint if that's the call you actually need.
Firecrawl vs Apify vs ScrapingBee at scale
Three contenders, three axes that decide the pick. Prices are current as of 2026-07-01.
| Firecrawl | Apify | ScrapingBee | |
|---|---|---|---|
| LLM-output fit | Markdown, deterministicJson, and grounded answer formats native (release notes) | Raw HTML or per-actor JSON; you convert for the model | HTML and JSON; no Markdown-for-LLM primitive (scrapingbee.com/pricing) |
| ~100K pages/mo | Standard, $83/mo yearly, 100K credits (pricing) | $199 Scale tier + per-actor compute-unit metering (apify.com/pricing) | $49 Freelance (250K credits) but JS + premium proxies cost multiple credits per call (pricing) |
| Agentic features | /monitor, Research Index, Agent endpoint, keyless MCP (changelog) | Actor marketplace, schedules; no LLM-native search | Proxy rotation, JS rendering; no agent-native tooling |
LLM-output fit: Firecrawl wins, and it isn't close. Markdown-by-default plus deterministicJson means the model on the other end gets ingestible text without an HTML-to-text pass.
Price at 100K pages a month: ScrapingBee's headline $49 Freelance tier (250,000 API credits) looks cheapest, but that's the trap (scrapingbee.com/pricing, fetched 2026-07-01). ScrapingBee credits are not 1:1 with pages; JavaScript rendering and premium proxies multiply the credit cost per call, so a JS-heavy 100K-page workload can chew through 250,000 API credits faster than the number implies. On plain, static pages ScrapingBee edges the price. On the JS-heavy pages agents actually hit, Firecrawl's flat 1-credit-per-page on Standard is the more predictable spend. Call it a split, with the nod to Firecrawl for anything modern-web.
Agentic features: Firecrawl wins outright. No competitor ships a change-monitor, a research index, and a keyless MCP an agent can call with zero setup.
The verdict, updated
The rating moves to 4.5 from 4.4. One sentence on why: the July changes made the same-priced product do more, keyless drops the first-run friction that used to cost a demo, and deterministicJson makes extraction both cheaper and repeatable, so the value-per-dollar rose without the dollar rising. The candid cons haven't gone anywhere. Crawl credit math still bites, Agent pricing is still dynamic and hard to budget, and the Research Index only pays off if you're building over arXiv.
Here is the pick by who you are.
Solo builder. Start keyless or on the free tier to prove the pipeline, then move to Hobby at $16/mo yearly the moment you need real concurrency or more than 5,000 credits (firecrawl.dev/pricing, fetched 2026-07-01). That is the whole ladder for one person shipping a side project or a niche tool. Don't overbuy Standard until your monthly burn actually crosses ~6,000 credits.
Small AI team. Standard at $83/mo yearly is the tier, full stop, at 100,000 credits and 50 concurrent requests per firecrawl.dev/pricing (fetched 2026-07-01). You get room to run /monitor and the Agent endpoint without watching the meter. Add the Research Index only if your product searches AI/ML literature; otherwise it's a feature you'll never call.
Enterprise. Self-host the AGPL-3.0 core if you have the DevOps to run browser pools and proxies, or contact sales for the zero-data-retention, SSO, and SLA package. At very high volume the self-host unit economics get interesting, which is the same crossover we flagged in May.
Against the field: Firecrawl versus Apify and Browse AI still lands the same way for LLM-destined data. Firecrawl is the pick when the output feeds a model. Apify wins when you need a configurable actor marketplace, and ScrapingBee wins on price for plain static pages at low complexity. For an AI agent that has to read the open web and hand it to an LLM, Firecrawl in July 2026 is the tool to buy, and Standard is the tier most teams should land on.
Ready to try it?
Try firecrawl →