Skip to main content
Private access is opening up — request an invite
askFinz
Web index · for AI and LLMs

Most AI reads arented search result.This one reads the web.

Giving a model web access usually means paying someone to pass your question to a search engine and hand back what it returned. askFinz reads pages itself, keeps what they mean, and files each one as the kind of document it actually is — 29 collections, answered from a standing index rather than fetched while you wait.

12 collectionsindex/

research-and-data/works
news-and-briefings/articles
encyclopedia/entries
industry-and-sector/documents
markets-and-filings/filings
medicine-and-clinical-research/records
products-and-reviews/products
property-and-construction/listings
trademarks/marks
code-repositories/repositories
books-and-manuals/documents
course-material/courses
filed by what it isa page each
12 of 29 named here
The short version
  • askFinz runs its own crawler and its own index — it reads pages first-hand rather than passing your question to a search engine.
  • A call returns the matched passage and the address it came from, not a ranked list you then have to fetch and read.
  • Every page is filed by what the document actually is — a judgment, a standard, a filing — across 29 collections, rather than as undifferentiated web pages.
  • $1.00 per 1,000 searches bundled, $0.75 as a top-up, $3.50 standalone. One provider matches that rate on its own index; the table below names them.
  • Buy something else if you already know the URL and just need it read cleanly, or if ranked links with no page text are genuinely all you need. Both cases are argued further down, with links.
Why this one

Four things that are true here and rare elsewhere.

None of these is a claim about being biggest or fastest. They are claims about shape — what gets read, how it is kept, and what comes back — and each one is checkable on a page of this site.

Filed by what it is, not where it was found

A judgment, a standard, a filing and a clinical study are not the same kind of document, and flattening them into "web pages" throws away the structure that makes each one checkable. askFinz keeps each as its own kind of record, with the identifiers, jurisdiction and dates attached.

See every collection

Read ahead of time, not fetched on demand

A relay passes your question to a search engine and hands back what it returned, which means it can only reach what that engine chose to show. askFinz reads pages continuously and holds what they mean, so the answer is a lookup rather than a live fetch — and it can reach material no engine surfaced.

How the index works

The passage comes back with the answer

Not a link to go and read, and not a summary you have to trust. The matched passage arrives with every result, and every record keeps the address it was read from, so any answer can be followed back to the publisher.

What the index holds

A platform, not an endpoint

Search, research, chat, code, news and the rest run on the same index, with one sign-in. That is the part no search API sells: a rate card gives you results to build on, not a place to do the work.

Every workspace
What changed, and when

Not our opinion. The record.

Owning the index used to be an eccentric thing to do. Over about eighteen months the case for it stopped being an argument and started being a series of events, none of them ours. Here they are, oldest first, with who reported each one.

  1. August 2025

    The budget option disappears

    Microsoft retires the Bing Search API with limited notice, breaking products built on it. Much of this market ran on that endpoint, and its replacement is a different product at a considerably higher price — which is a large part of why web access costs what it does today.

    Reported by Firecrawl and by serp.fast

  2. September 2025

    Cloudflare names re-crawling as the waste

    Announcing an AI index of its own, Cloudflare writes that for AI builders “crawling and recrawling unstructured content is costly, wastes resources, especially when you don’t know the quality and cost in advance”, and enumerates the stack it costs: compute, storage, databases, embeddings, chunking and models. Its answer is an index sites opt into — so its coverage is whatever sites choose to join, which is a real difference from reading the open web.

    Cloudflare blog, 26 September 2025

  3. December 2025 – July 2026

    Relaying an engine turns out to carry legal risk

    Google sues SerpApi over DMCA claims. A court dismisses the core claims in July 2026, with leave to amend part of them. The outcome matters less than what it exposes: a business built on passing queries to somebody else’s engine has a dependency that can be litigated, and it is not the operator’s to settle.

    Reported by serp.fast

  4. April 2026

    The list of independent indexes gets shorter

    Stract is archived read-only; Gigablast has been offline since 2023. Of the handful of search indexes built from an operator’s own crawl, two of the named ones are now gone. Running one is not merely rare — it is getting rarer.

    Reported by serp.fast

  5. August 2026

    Amazon ships an index of its own

    Bedrock Web Search launches on “a web index operated by Amazon, spanning tens of billions of documents refreshed continually”. Anyone claiming that everybody else rents is now wrong; the true and sufficient claim is that most AI products do.

    AWS product documentation

  6. September 2026

    OpenAI turns out to have been building vertical indexes since 2023

    An investigation drawing on antitrust testimony, job postings and an observed result-source field concludes that OpenAI runs an index of its own and that it is not one index but a family of vertical ones — general web, PDF, YouTube, news, arXiv, Wikipedia, local, finance, legal, medical, shopping and images. That is the argument this page makes in its first pillar, reached independently and at a scale nobody here can match. Flattening the web into undifferentiated pages is what the largest AI company stopped doing.

    Peec AI, 4 September 2026

The last entry is the one worth sitting with. askFinz’s first claim is that the web should be filed by what each document is rather than flattened into pages, and the largest AI company in the world appears to have reached the same conclusion on its own, with resources nobody here can match. What that looks like here is 29 collections you can open and read.

How it is read

Many small machines, not one rented service.

Reading the web continuously is the expensive part of running an index, and it is the reason most products rent their results instead. askFinz spreads the work across pools of ordinary machines — a spare desktop, a container on a server already running, or a dedicated box booted from an image — each joining a pool and taking work when it has capacity.

the reading fleetpool shapes are real; node counts here are illustrative
wsl/a spare desktop, one command
docker/a container on a server you already run
sealed/a dedicated box, booted from an image
Pages are read by whichever pool has capacity, then filed by what the document is. Adding capacity means adding a machine to a pool, not renting a larger plan — which is why the cost of reading more of the web rises with hardware you already own rather than with someone else’s rate card.

Compute follows the work. A pool takes what it can handle and no more, so a busy hour spreads across more machines rather than queueing behind a fixed plan.

Storage is what a page means. What is kept is the meaning and the address it came from, not a copy of the web — which is why a stored page costs a fraction of what the original weighed.

Read under our own name. Requests identify themselves and are answerable to a published crawler policy — no rented proxy pool standing between us and the site being read.

The node programme is open — lend a machine, run it as a container , or boot a dedicated box from an image. Why we do not use a rented proxy pool is set out in how askFinz reaches a site.

What it costs

At the floor, and honest about it.

Every rate below was read off the vendor’s own pricing page, and every capability off their own documentation, on 7 September 2026. The table is grouped by the thing that matters most and is hardest to see from a rate card: whether a provider holds an index of its own, borrows one, relays a search engine, fetches a page you already named, hands you an archive to process, or meters the web inside a model. Cheaper rows are in here rather than hidden underneath — most of them sell a call that never returns the words on the page, which is a column now instead of a footnote.

Capabilities from each vendor’s own documentation, rates from their own pricing page, both checked 7 September 2026. Rates are the cheapest published self-serve figure per 1,000; where a vendor publishes none, the cell says what they price by instead rather than carrying an estimate. Free tiers are excluded throughout — they are a trial, not a price. Rows run cheapest first within each group.
ProviderWhat one call returnsPage textFiled by document typeWorkspaces on topFrom, per 1,000vs askFinz bundled
Holds its own indexCrawls the web itself and answers from what it holds. The only group that can decide what is in the index.
askFinzThe matched passage, with its sourceIncluded$0.75top-up; $1.00 bundled, $3.50 standalone
Parallel
Ranked URLs and compressed excerptsIncludedcompressed, not the full page$1.00turbo and fast; covers the FIRST 10 RESULTS, then $1.00 per 1,000 further results. Basic and advanced $5.00about the sameat 10 results a call. A call returning 100 costs about $91 per 1,000 on the same published rates
Keiro
Ranked results, with page text at the higher tierIncluded3 credits; links alone cost 1$2.40best tier; $4.00 and $7.20 below it. Links only $0.802.4× more
Keenable
Ranked results with long snippetsBilled extralong snippets, not the page$4.00pay as you go; $1.00 only on dedicated capacity4.0× more
Yep
Ranked results with query-relevant highlightsBilled extrahighlights around the match$4.00advanced tier $8.004.0× more
Brave Search API
Ranked results and descriptionsNot returned$5.00answers tier $4.00 plus tokens5.0× more
Exa
Ranked results; page contents billed per pageBilled extra$1.00 per 1,000 pages on top$5.00answer tier; search $7.00, deep $12.005.0× more
Linkup
Results with content, or a cited answerIncluded$5.00standard depth; deep search roughly ten times this5.0× more
Amazon Bedrock
Semantic passages, inside BedrockIncluded$7.00Amazon's own index, launched August 20267.0× more
Partly its ownRuns some crawling of its own and leans on an outside engine for the rest.
You.com
Ranked results; full page content billed per pageBilled extracontents $1.00 per 1,000 pages$5.001–100 results per call5.0× more
Perplexity
Raw results with advanced filteringNot returned$5.00Search API, per 1,000 requests5.0× more
Google Vertex AI Search
Results from Google's index, groundedIncludednot per searchper query PLUS per GB indexed — no single per-1,000 figure
Relays a search engineHolds no index. Passes your query to an engine and hands back what it returned — cheap, and genuinely useful, but a different product.
DataForSEO
A search engine's results — no page textNot returned$0.60queued mode; live calls $2.00cheaper, thinner call
SearchAPI
A search engine's results — no page textNot returned$1.00at volume; $4.00 at the entry planabout the same
Bright Data
A search engine's results, at scaleNot returned$1.30SERP API; $1.50 without a commitment1.3× more
SerpApi
A search engine's results — no page textNot returned$1.97at the largest published tier; $25.00 entry2.0× more
Tavily
Aggregated sources, content and a short answerIncluded$5.00largest plan; $8.00 pay as you go, advanced double5.0× more
Serper
A search engine's results — no page textNot returnednot per searchno public per-1,000 rate published
Fetches a page you nameNo search at all. You already know the URL; it reads that one page cleanly.
Firecrawl
One named URL, as clean markdownIncludednot per searchpriced per page fetched, not per search
Apify
Whatever the scraper you chose was built to takeIncludednot per searchpriced per run and per result, not per search
A bulk archive you process yourselfNo queries and no ranking. A large snapshot of the web that you store, parse and search on your own hardware.
Common Crawl
A monthly archive of raw pages, to process yourselfIncludedraw, unparsed, months oldnot per searchfree to download; you pay in storage and compute
Web access billed inside a modelNot an index you can query — the model's own access to one, metered per call by the model provider.
xAI web search
Results handed to the model, inside the completionNot returned$5.00per 1,000 web search tool calls5.0× more
Google grounding
A grounded answer, not an index you can queryNot returned$14.00per 1,000 grounded queries14.0× more

Read the last three columns if the rates look alike. Half the market’s cheapest rows do not return the words on the page at all, and among those that do, only askFinz files what it reads by what the document actually is — a judgment, a standard, a filing — and only askFinz puts workspaces on the index rather than handing back an endpoint. Several rows beat askFinz on something: Parallel matches the price, Firecrawl reads a named URL askFinz will not, the relays are far cheaper for ranked links, and Common Crawl costs nothing if you own the hardware to process it. Each of those pages says so.

You will find cheaper rates than these. A separate group of services — SerpApi, DataForSEO, Serper, SearchAPI and others — quotes from about $0.60 per thousand, and they are worth understanding rather than dismissing: they don't hold an index of their own, they pass your query to a search engine and hand back what it returned. That is a genuinely useful thing to buy, and it is priced accordingly. It is also a different product from this one. They return the search engine's results — titles, links and snippets — not the text of the pages behind them, so they can only reach what the engine chose to show, and none of the material we hold because we read it from the source. Among services that run their own index, the rates above are the market.

What the model providers charge for web access. If you are giving an LLM the web rather than querying an index yourself, the meter runs inside the model provider: xAI’s web search tool is $5.00 per 1,000 calls, Perplexity’s Search API $5.00 per 1,000 requests for raw results with filtering, and Google’s grounding $14.00 per 1,000 grounded queries. They are not in the table above because they are not an index you can query — they are the model’s access to one, and what comes back is results or a grounded answer rather than the passage and its source. Rates confirmed 7 September 2026.

The budget option is gone. Bing's Search API — what much of this market was built on — was retired in August 2025. Its replacement is a different kind of product at a considerably higher price, which is part of why search costs what it does elsewhere.

Rates read from each provider's own pricing page on 7 September 2026. Where a provider gets cheaper with volume, the comparison uses their best rate, not their entry price. Rates are matched on what the call returns: askFinz search returns the relevant passage from the page with every result, so where a provider charges separately for page text, that is the rate shown. Providers change pricing — check theirs before relying on this.

How to choose

Five questions that settle it faster than a rate card.

  1. 01

    Do you need the page, or a link to it?

    The cheapest services hand back a search engine's ranked links. If your model then has to fetch and read each one, that fetch is the real cost — and it is not on the rate card you compared.

    Compare the rates, and what each call returns
  2. 02

    Does it matter what kind of document it is?

    For general questions, no. For a judgment, a standard, a filing or a clinical study it decides everything: those carry identifiers, jurisdiction and status that a generic page index throws away.

    See every collection and how it is filed
  3. 03

    Who owns what you are searching?

    A provider relaying an engine can only reach what that engine chose to show, and their price moves when the engine's does. A provider with its own index sets both.

    Owning an index versus renting one
  4. 04

    Is this a call in a loop, or work someone does?

    If you are building an agent framework, an API with good integrations may save more in engineering time than it costs in rate. If people are doing the work, a workspace they can open beats an endpoint.

    The workspaces built on the index
  5. 05

    What happens when the answer is questioned?

    Ask whether you can get back to the source. An answer with the passage and its address survives being challenged; a summary of a summary does not.

    How the index works, end to end
What it costs us

Affordable to run is why it is affordable to buy.

A search index has three running costs: reaching pages, understanding them, and keeping what you learned. Rent any of the three and the price of your product is set by somebody else’s margin — which is why so many AI products charge per web call, and why those calls cost what they do.

askFinz owns all three. Reading spreads across pools of ordinary machines rather than reserved capacity billed by the request, so capacity grows by adding a machine instead of moving up a plan. Understanding a page runs on the hardware in those machines, bounded by what the fleet already has rather than by a per-token bill. And what is stored is what a page means plus the address it came from, not a copy of the page — a fraction of what the original weighed.

None of that is charity, and none of it makes askFinz cheapest at any price: the matrix above shows one provider matching the bundled rate. What it does mean is that the rate is set by hardware we can add to, not by a contract we have to renew.

Compute you add to, not rent

A pool takes what it can handle and no more. Reading more of the web means another machine in a pool, so cost tracks hardware rather than a plan that has to be upgraded.

Storage of meaning, not of pages

What is kept is what a page means plus the address it was read from. That is a small fraction of what the page weighed, and it is what makes holding a standing index affordable at all.

Reading done once, reused forever

A page is read when it is found and re-read when it changes — not once per question. Every question after the first is a lookup, which is why the rate does not rise with how often you ask.

The point of keeping it cheap to run is that the people who most need an index rarely have a budget for one — a researcher, a small firm, a team automating something nobody has automated before. A rate that tracks hardware rather than a rented plan is what makes building on the web affordable for them, which is the argument for owning the thing rather than renting it. Owning an index versus renting one sets that out in full.

Being straight about it

When something else is the better buy.

You need to fetch a named URL

askFinz answers from what it has already read. If your job is “give me this page, now”, a scraping service is the right tool. Compared with Firecrawl · with Apify

You want ranked links, as cheaply as possible

Relay services start well below anything here, because they hand back a search engine’s results rather than the pages behind them. What SERP APIs are · compared with SerpApi

You are building an agent framework

Several search APIs are built to be called from inside a loop you own, and their integrations save real engineering time. Compared with Tavily · with Exa

Every comparison on this site is written to survive the other company reading it. Rates come from their own pages, capabilities from their own documentation, and where a rival is cheaper or does something askFinz does not, the page says so. The full set is at compare, and the two category explainers are owning an index versus renting one and indexing versus scraping.

Questions

The ones worth asking first.

Is askFinz the cheapest web index?
It is at the floor rather than below it, and the honest answer depends on what the call returns. Among services that hold their own index and return the matched passage, askFinz's bundled $1.00 per 1,000 is the cheapest published rate, tied with one other. Services that return ranked links without the page text are cheaper — that is a thinner call, and a fair comparison says so. The full table, with every rate read off the vendor's own page, is on the pricing page.
What makes this different from a search API?
A search API gives your code results to reason over. askFinz gives an answer with the passage it came from, on an index filed by document type, inside workspaces you can actually work in. If what you want is an endpoint to build against, several of the services compared here are a better fit and the comparison pages say which.
Does askFinz rent its search results from a search engine?
No. The index is read and held by askFinz. That is unusual but not unique — a small number of others run their own index too, and the comparison pages name them rather than implying otherwise.
How does the reading fleet keep costs down?
Reading is spread across pools of ordinary machines rather than concentrated in rented capacity billed per request. Adding capacity means adding a machine to a pool, so the cost of reading more of the web tracks hardware rather than someone else's rate card. That is the mechanism behind the published rate.
Can I use askFinz for crawling or scraping?
Not directly. Scraping and crawling services take a URL you name and return its content; askFinz answers a question from material it has already read and filed. If your job is to fetch specific pages on demand, the crawling comparison explains which tools do that and why askFinz is the wrong answer for it.
Read it yourself

Open a collection and see what is actually in it.

Every collection page publishes its own size, how it divides, and what one record holds — read from the running system rather than written down.

The best web index for AI and LLM workflows — askFinz