Firecrawl now charges only when a scrape returns a page, and Enhanced mode costs nothing extra. The "8x cheaper" agent claim can't be checke

What we checked

• Does the cluster show a broader market or architecture shift, not just one price change? partly confirmed. The cuts fit a wider move. Firecrawl is making the parts it pays for by usage (web pages, research papers, the free Research Index) free or cheaper, while it builds up its data-access business: the Alexandria data-provider catalog, its plugins for Claude, ChatGPT and Codex, and a $75M Series B. Whether this is a market-wide shift can't be shown from one vendor's evidence. [7][8]
• Are failed requests now free, and what counts as 'failed'? partly confirmed. Yes, if 'failed' means Firecrawl returned no document. Those requests cost 0 credits. A 403 or 404 page that Firecrawl does return still costs 1 credit, and some add-ons charge even when the scrape fails: threat scans (2 credits), the prompt-injection check (5 credits), lockdown cache misses (1 credit), and browser minutes. [1][2]
• Is Enhanced mode on /scrape now free of any surcharge? confirmed. The docs say enhanced proxy requests cost the same 1 credit as basic ones, and an automatic retry on an enhanced proxy isn't charged separately. A review last edited in June 2026 listed Enhanced/Stealth at 5 credits per page. [3][1][4]
• Is /agent 8x cheaper? unresolved. Agent pricing is still 'dynamic', based on agent tokens. Firecrawl doesn't publish a token-to-credit rate, so there's no published before-or-after price to test the 8x against. The docs do show the agent now runs only on Spark 2, which they call cheaper than Spark 1, and that failed runs are no longer billed. [6][2]
• Is the Research Index free, with 43M+ papers? partly confirmed. The pricing page lists the paper endpoints as free in every category. Firecrawl's changelog entry for Aug. 13 cites 41M+ papers, not 43M+. The 43M figure appears in the X post and in a third-party review, but not in the Firecrawl pages I opened. [2][7][9]

What we found

• Billing now depends on whether Firecrawl returned a document, not on the HTTP status the target site sent back. A request where Firecrawl returned no document costs 0 credits. A 403 or 404 page that does come back costs 1 credit. Some features charge even when the scrape fails: threat scans (2 credits per URL), the prompt-injection check (5 credits once it has run), and lockdown cache misses (1 credit). [1][2]
• Enhanced proxies now cost the same 1 credit as a basic request, and an automatic retry on an enhanced proxy isn't billed separately. The automatic upgrade happens only when the site returns 401, 403 or 429, and at most once per request. For comparison, a review last edited in June 2026 listed Enhanced/Stealth at 5 credits per page. Proxidize reports that the changelog dates the end of that charge to Aug. 17, 2026, but I didn't see that entry on the changelog page I fetched. [3][4][5]
• /agent is still a preview feature with 'dynamic' pricing. It includes 5 free daily runs, after which it is charged in credits based on agent tokens, at a rate Firecrawl doesn't publish. The docs say most runs use 'a few hundred credits.' maxCredits defaults to 2,500, every run now uses spark-2 (Spark 1 models are deprecated), and failed runs, including those that hit the credit limit, aren't billed. Before the change, the June 2026 review said agent runs were charged whether or not they succeeded. [6][1][4]
• The Research Index paper endpoints (search_papers, related_papers, inspect_paper, read_paper) are listed as free in every category, including AI/ML and life sciences. Firecrawl's changelog entry of Aug. 13 already said the whole index was free and cited 41M+ papers. So the 'free index' part of the September announcement seems to formalize an earlier change rather than introduce one. [2][7]
• Some per-page charges haven't changed. JSON, Question and Highlight formats still add 4 credits per page, search costs 2 credits per 10 results, and Interact costs 2 credits per browser minute. Plan prices don't appear to have fallen: annual-billed rates are $16, $83, $333 and $599 a month, and an independent check on Sept. 7 recorded monthly prices of $19, $99, $399 and $749. [8][9]
• The price cuts came alongside a broader product push in the changelog: the Developer Index (Aug. 20), Alexandria and a $75M Series B (Sept. 22), plugins for Claude/Claude Code and ChatGPT/Codex (Sept. 24 and 29), and Universal Scrape across pages, files and data providers (Oct. 8). In those plugins, finding and inspecting providers is free, while running them is billed at each provider's listed price, according to a summary I couldn't confirm on the page itself. [7]

What it means

For scraping pipelines, two of the changes are real and easy to measure. A blocked fetch that returns nothing costs nothing, and retrying it with Enhanced proxies no longer costs five times as much. You can stop maintaining your own fallback to stealth proxies and let the default auto setting handle it. Keep a few cases in your cost model, though: 403 and 404 pages that do come back still cost a credit, and the prompt-injection check and threat scans charge even when the scrape fails. Don't treat "failed is free" as a blanket rule for blind retries. A retry that brings back an error page is a billed success.

For /agent, treat "8x cheaper" as a claim from the vendor, not a rate you can plan around. Pricing is still per token and the conversion rate isn't published. Before you move more research work onto it, set maxCredits explicitly, log creditsUsed per run, and compare those numbers against your own history from before September. Failed runs and runs that hit the credit limit are no longer billed, so a hard cap now limits what you lose on a bad run. More broadly, Firecrawl is making raw retrieval cheaper and charging for extraction formats, add-on checks and data from its provider catalog. Budget for those charges, not just the per-page rate.

Still open

• What is the token-to-credit rate for /agent, and what baseline (Spark 1? a specific credit amount per run?) is the '8x cheaper' claim measured against?
• What is the exact date of the announcement and of each change? The X post couldn't be retrieved, and Proxidize cites an Aug. 17 changelog date for Enhanced mode that I didn't see on the changelog page I fetched.
• Does the Research Index have 41M+ papers (changelog) or 43M+ (announcement)?
• Do these changes also apply to the free tier and to zero-data-retention (ZDR) runs, where agent AI credits are billed at 2x?