Scry · Comparison

Scry, next to the search APIs

Ten products agents reach for when they need the internet, verified against their live pages with the date on every claim: the survey you would otherwise run yourself. They sell ranked results and finished answers over closed indexes. Scry sells the layer underneath: the corpus itself, open to SQL.

The market, one row per product

ProductClassA call returnsIndex & corpusResult depth & limitsPriceLatencyMCPStatus / SLABackingVerified
Exaranked web searchranked links + extracted contentown neural index (transformer embeddings, own GPU fleet); claims 500B URLs tracked, hourly updates; no document count published10 included; each further result billed$7 /1k searches + $1 /1k extra results; contents $1 /1k pages; deep tiers $12–15 /1kclaims ~250 ms (instant tier) to ~40 s (deep-reasoning)yesstatus page 99.97%; no SLA published$335M+ raised; $2.2B valuation (May 2026)2026-09-03
Brave Search APIranked web searchranked links, news, images; answers tierown index: claims 30B+ pages, 100M+ updates/day, fed partly by opt-in browser Web Discovery datatop-k pages per call, 50 qps$5 /1k requests; answers $4 /1k + $5 /1M tokensSOC 2 Type II; no SLA % publishedBrave Software (browser-scale org)2026-09-03
Tavilyranked web search≤20 results as ≤500-char snippets; optional answerhybrid: own crawler (presents as Googlebot-compatible) + per-query aggregation of third-party search data; no size published20 results, 500-char chunks — hard shape$0.008 /credit; search 1–2 credits; plans from $30/mo (4k credits); free 1k/moclaims 180 ms p50yesclaims 99.99% SLA; public status page$25M raised (Series A, Aug 2025)2026-09-03
Parallelranked web searchranked results; task tier returns finished research jobsown crawl + index; no size published; vendor-run benchmark tables at parallel.ai/benchmarks10 results per search requestsearch $1–5 /1k; extract $1 /1k results; task processors $5–2,400 /1k; free 5k/moyesSOC 2; enterprise ZDRfounded by Parag Agrawal (ex-Twitter CEO)2026-09-03
Linkupranked web searchcited answers + full-text snippetsaggregated web retrieval; sources and freshness tunable per call; index provenance unpublishedanswer + snippets per callsearch $5–6 /1k; fetch $1–6 /1k; research runs $0.25–2.50 each; free 4k queriesclaims 1–3 s sync search, <2 s fetchSOC 2 Type II on all plansclients incl. McKinsey, Databricks, Cohere2026-09-03
Perplexity Sonaranswers & reportsprose answer + citations; raw search endpointown crawler + index built for the consumer answer engineone answer per callsearch $5 /1k; Sonar $1/$1 per 1M tokens + $5–12 /1k requests; Pro $3/$15 + $6–14; deep research token-billed + $5 /1k searchesPerplexity (consumer-scale org)2026-09-03
Elicitanswers & reportspapers, screening tables, extraction columns, reports138M papers + 545k trials — Semantic Scholar, PubMed, OpenAlex merged; the moat is the workflow, not the retrieval300 results/request (Pro) to 10,000 (Enterprise); 100 req/min per IP on every planAPI from the $49/mo Pro plan; Scale $169/mo; Enterprise customyes (OAuth)~$31M raised; ~57 people (trackers); PBC2026-09-03
Consensusanswers & reportstop-20 papers with extracted study-design metadatascientific papers; provenance not detailed publicly20 results per call — the whole surfacefrom $0.10 /call + platform fee; application-gated, custom quotesno2026-09-03
Semantic Scholaropen scholarly datascholarly metadata records; citation graph214M papers, 2.49B citations, 79M authorspaginated; 1 request/sec keyedfreebest-effort (rate-limited for politeness)Ai2 (Allen Institute for AI)2026-09-03
OpenAlexopen scholarly datascholarly metadata records; group-by aggregations323.5M works (live API count on the verified date), CC0paginated; full snapshot downloadablefree; paid premium for faster update feedsbest-effortOurResearch (open infrastructure)2026-09-03
Scryprogrammatic searchthe rows you asked for — or the answer computed from all of themown archive of the public internet as one corpus; coverage published as measured dents (26.9B Reddit comments = 94% of all ever written)no page 2 — your LIMIT; aggregations read every matching row (a published study read 50.46B rows in 15 m 46 s)free under slack; congestion reprices; each response reports its costengine-reported per query; a 111-iteration recursive loop returns in ~50 msyesprivate alpha, own hardwareExoPriors2026-09-03

Latency, index-size, and SLA figures are each vendor’s own published claims. A blank cell means the fact could not be verified either way.

Capabilities

CapabilityScryExaBraveTavilyParallelLinkupSonarElicitConsensusS2OpenAlex
Live-web retrieval on demand
Archive depth (years of record, stable refs)
Full-text content extraction
SQL: aggregation, joins, GROUP BY
Recursion & graph programs
Computation without a result cap
Corpus enumerable & auditable
Per-row provenance fields
Bulk snapshot export
Synthesis reports / review workflow
MCP server

● yes   ◐ partial   — no   blank: not verified.

Which one, when

The ten best links from the live web, in under a secondExa or Brave
One cited answer with no agent scaffoldingPerplexity Sonar
Drop-in web tool call for a LangChain-era agent frameworkTavily
A whole research task as one metered requestParallel (task tier)
A systematic review: screening, extraction, PRISMA-shaped outputElicit
Scholarly metadata in bulk, free, no stringsOpenAlex or Semantic Scholar
Count, join, rank, sweep, or walk the whole record — not the top tenScry

Where each wins, where each stops

ProductWinsStops
Exastrongest semantic top-k over the live web; richest latency ladder; heaviest-funded pure search APIranked lists only; index unauditable; results past ten billed per item
Brave Search APIlargest independent index for rent; cleanest per-query price in the ranked classtop-k pages; the index cannot be queried as data
Tavilybroadest agent-framework integrations; snippets pre-cut to context windows20 results × 500 chars is the product’s hard shape
Parallellowest price floor ($1 /1k); task tier runs whole jobs per requestranked results and finished outputs; no computation over the index
Linkupcited-answer ergonomics; per-call source control; low priceretrieval underneath is a black box you steer but cannot inspect
Perplexity Sonarone call replaces a search-read-synthesize loopprose and citations only, never the corpus; costs compound on tokens and requests at once
Elicitreal systematic-review machinery no raw corpus providesscholarly records only; hard caps; per-IP limit that paying more (below Enterprise) does not lift
Consensusclean extracted study-design metadata per papertwenty results per call is the entire programmatic surface
Semantic Scholarfree citation-graph truth; bulk downloadsmetadata only; 1 req/sec; all computation happens on your machine
OpenAlexthe whole catalog, downloadable, CC0 — Scry serves OpenAlex-derived relations itselfserver-side power ends at filtered listing and group-by

The Scry row, expanded

Query surfaceread-only SQL (ClickHouse dialect) + WITH RECURSIVE + fixpoint graph programs; vector helpers; rerank
Corpusthe public internet as one corpus: forums, scholarly literature, prediction markets, social archives, newsletters, public records — billions of new posts and comments daily
Depthno page 2; answers computed over every matching row (published study: 50.46B rows in 15 m 46 s)
Coverage honestypublished as measured dents — 26.9B Reddit comments, 94% of all ever written, the miss computed by exact subtraction
Transparencylive schema is the contract; every response returns the executed SQL and engine accounting
Provenancesource, batch, and timestamp fields on every row
Connectone MCP URL (mcp.scry.io) for Claude Code, Codex, ChatGPT, Cursor; HTTP API with the same key
Pricefree under slack; under measured congestion, admission reprices; each response reports its cost
Docsturing-complete search and the calculus it implements, at /docs

Sources for the Scry figures: studies carry the engine-reported timings, and the docs carry the query calculus. The asymmetry the whole page reduces to: a ranked list can always be built on top of a corpus, but a corpus cannot be recovered from ten links.

What they do better, today

Ranked one-shot answering over the live web is their whole product, and the leaders are excellent at it: when your agent needs the ten best links or one cited answer in two seconds, use them. Their extract and crawl endpoints reach arbitrary URLs the moment you ask. Elicit’s screening and extraction tables are years of dedicated workflow product. And they carry the operational trappings of bigger teams: SOC 2 attestations, public status pages, enterprise SLAs. Scry is a private alpha on its own hardware, and this page will keep saying so for exactly as long as it stays true.

FAQ

Which of these should my agent actually use?
For ten good links or a one-shot cited answer over the live web in a couple of seconds, use a ranked search API — Exa and Brave are the strongest for retrieval, Sonar for finished answers. For a systematic literature review with screening and extraction, Elicit sells exactly that workflow. When the question is computational — count, join, rank, sweep, walk a graph, check every row rather than the top ten — that is the layer Scry serves.
Is Scry a web search API?
Scry serves search and fetch tools over its own archive of the public internet, so MCP clients can use it like one. The difference is what sits underneath: the same corpus answers arbitrary read-only SQL, so retrieval is one query shape among many rather than the whole product.
Why does the Scry row show no result cap?
Because the unit is the query, not the result list. A statement carries its own LIMIT, and aggregations return computed answers over every matching row. One published study read 50.46 billion rows in under 16 minutes.
What do these products do better than Scry?
Ranked one-shot answering over the live web — that is their entire product, and they are excellent at it. Fetching an arbitrary URL this second. Elicit’s screening and extraction workflow. And operational maturity: larger teams, SOC 2 attestations, public status pages with long uptime histories. Scry is a private alpha run on its own hardware.
How current is this page?
Every row carries the date its facts were verified against the vendor’s live pages, and prices are quoted from those pages, not from memory. A blank cell means we could not verify the fact either way. When a number drifts, write [email protected] and we will re-verify.

Create an account to run the comparison yourself — read-only SQL over the corpus, provenance on every row.

All product names are trademarks of their respective owners. Prices and claims are quoted from each vendor’s public pages on the verification dates shown; vendor-published performance numbers are labeled as such.