Parallel API review
Best for per-request research you can price in advance
POST https://api.parallel.ai/v1/search x-api-key: $PARALLEL_API_KEY # default mode: advanced, $5 per 1,000, about 3 seconds # "mode": "fast" $1 per 1,000, about 700ms # "mode": "turbo" $1 per 1,000, about 200ms
The Parallel API is a set of web APIs built for agents. Search returns ranked URLs with compressed excerpts, Extract returns page contents, and Task runs deep research with structured output. FindAll builds a list of entities that match a description, and Monitor watches the web for changes.
The common thread is the pricing. Parallel charges per request, never per token, so it says you always know the exact cost of a query before you run it.
For a GTM team, two of those APIs are the point. Search is the agent search tool, competing with Exa, Tavily and Perplexity. Task is closer to enrichment: Parallel lists database enrichment and structured enrichments among what it is for, priced by the processor you choose (Parallel's name for a tier of work, from Lite up to Ultra) rather than by what comes back.
The finding worth knowing before the first call is the default. A Search request with no mode set runs in advanced mode, at $5 per 1,000. Setting fast or turbo mode costs $1 per 1,000. One parameter is a factor of five on every search.
The rest of the category is in the Sourcing APIs listing.
What a request costs, by processor
Every Parallel API has processors or modes, and the processor sets the price. Search is cheap at any setting; Task spans a factor of 480 between its lightest and heaviest processor.
| API and processor | Per 1,000 requests | Latency | What it is for |
|---|---|---|---|
| Search, turbo | $1 | ~200ms | High-volume lookups. |
| Search, fast | $1 | ~700ms | Interactive agents. |
| Search, basic | $5 | ~1s | Longer excerpts per source. |
| Search, advanced (default) | $5 | ~3s | Multi-hop background research. |
| Extract | $1 per 1,000 results | 1s to 3s cached | Page contents from a URL. |
| Task, Lite | $5 | 5s to 60s | Basic retrieval, about 2 fields. |
| Task, Core | $25 | 1 to 5 min | Complex research, about 10 fields. |
| Task, Ultra8x | $2,400 | 8 to 30 min | Deep research, about 25 fields. |
A Search price covers 10 results, and each additional 1,000 results is $1 on any mode. Task has more processors than the table shows, all priced per 1,000 requests: Base at $10 sits between Lite and Core, and Core2x at $50, Pro at $100, Ultra at $300, Ultra2x at $600 and Ultra4x at $1,200 sit above Core. FindAll charges $0.25 a run plus $0.03 a match on Base, rising to $10 a run plus $1 a match on Pro.
Parallel's quickstart says its basic request uses the default advanced mode, about 3 seconds at the highest quality, and that most agents should add "mode": "fast". Advanced and fast are $5 and $1 per 1,000. An agent that leaves the default in place pays five times what Parallel itself recommends for most of them.
Parallel API pricing, plan by plan
Parallel API pricing is pay as you go per request, with $5 of free credits every month and up to $80 at signup.
Free credits
- Up to 5,000 requests a month
- Plus up to $80 at signup
- Startups can apply for up to $250
Search
- Turbo and fast $1, basic and advanced $5
- 10 results included
- 600 requests a minute
Task
- Nine processors, Lite to Ultra8x
- Citations, reasoning and confidence
- 2,000 runs a minute
Enterprise
- Custom rate limits
- Dedicated onboarding and support
- Early access to new products
The $5 of monthly credits is 5,000 searches in fast or turbo mode or 1,000 in the default mode, which is the same factor of five again. Parallel puts the free allowance at up to 5,000 requests a month, and that figure assumes the $1 modes.
Because Parallel does not bill tokens, a long Task answer costs the same as a short one on the same processor. That makes the processor the only decision, and Parallel's own guidance is to flex compute by the complexity of the task: start on Lite or Base and move up only for the fields that need it.
Limits on creating work, not on reading it
Parallel API rate limits apply to POST requests that create new resources. Search and Extract allow 600 a minute, Tasks 2,000 a minute, Chat and Monitor 300 a minute, and FindAll 300 runs an hour. GET requests for results and status do not count.
| Product | Default quota | What counts |
|---|---|---|
| Search | 600/min | Each POST to /v1/search. |
| Extract | 600/min | Each POST to /v1/extract. |
| Tasks | 2,000/min | Each Task Run created, including each input in a Task Group. |
| Chat | 300/min | Each POST to /v1beta/chat/completions. |
| Monitor | 300/min | Each POST creating a monitor. |
| FindAll | 300/hour | Each FindAll run created. |
The Parallel API rate limits favor polling: checking a task's status is free against the quota, so a pipeline can submit Task Runs at up to 2,000 a minute and poll as often as it likes. In a Task Group, each input counts as a run, so a group of 500 rows is 500 against the limit.
How to create a Parallel API key
Create a key in the Parallel platform and send it in the x-api-key header. The Search MCP server works without one, at lower limits.
The REST APIs
One key, every API.
- Send x-api-key on each request.
- Use /v1/search and /v1/extract for new work.
- The /v1beta paths are for existing integrations only.
The Search MCP server
Hosted at search.parallel.ai/mcp.
- Free to use anonymously, at lower rate limits.
- Pass a Parallel API key for the full limits.
- Mode and domain settings can go in the URL.
An anonymous MCP server is the cheapest way to try Parallel inside an agent, because nothing is billed and nothing needs setting up. Move to a key the moment anything depends on it, because the anonymous limits are lower.
A first Parallel Search call
This is the request from Parallel's own Search quickstart, which uses the default advanced mode. Add a mode to change the price.
curl https://api.parallel.ai/v1/search \ -H "x-api-key: $PARALLEL_API_KEY" \ -d '{"objective": "Find latest information about Parallel Web Systems.", "search_queries": ["Parallel Web Systems products"]}' # add "mode": "fast" to drop it from $5 to $1 per 1,000
Where the Parallel API documentation lives
The Parallel API documentation lives at docs.parallel.ai, which publishes an llms.txt index and a full-text llms-full.txt for agents.
| Surface | Where | What it answers |
|---|---|---|
| Search quickstart | docs.parallel.ai/search/search-quickstart | The request, the modes and the SDKs. |
| Rate limits | docs.parallel.ai/getting-started/rate-limits | Every quota and what counts. |
| Search MCP | docs.parallel.ai/integrations/mcp/search-mcp | Hosted setup for Claude, Codex and others. |
| Pricing | parallel.ai/pricing | Every processor, mode and fee. |
Parallel also lists Task MCP and integrations on its developers page, and publishes a changelog and a status page.
The MCP page lists Parallel's MCP access alongside the others in this directory.
Parallel's own walkthrough of the API
Parallel publishes this on its own channel, so it is Parallel's account of Parallel, not ours. It is here because a rate card cannot show you the thing moving, and a minute or two of the product is faster than any paragraph on this page at telling you whether the shape of it fits yours.
YouTube · Parallel Web Systems Introducing the Parallel Search API 1:21 · Published November 2025Parallel's 81-second launch film for the Search API, which it describes as built from the ground up for AIs rather than for people. It is a launch, not a walkthrough: it predates the turbo mode and shows no request, so the modes and the prices on this page come from Parallel's pricing and docs pages rather than from the video.
Published 6 November 2025 on Parallel Web Systems, the channel Parallel's own pages link to, and read on 23 October 2026. Nothing in it is a source for a price or a limit on this page: every figure in the tables above comes from a Parallel page.
Strengths and limits
Parallel's pricing is the most predictable in this category, and its range is the widest.
What it does well
- Priced per request, so the cost is known before the call.
- Search from $1 per 1,000 in fast or turbo mode.
- $5 of free credits a month, plus up to $80 at signup.
- Polling and result reads do not count against the rate limit.
- A Search MCP server that works anonymously.
- Task and FindAll go past search into structured research and lists.
Where it falls short
- The default search mode is the $5 one.
- Task processors span $5 to $2,400 per 1,000, so the choice matters.
- Search is capped at 600 requests a minute by default.
- The anonymous MCP server runs at lower limits than a keyed one.
- No contact data, so it sits beside an enrichment provider rather than replacing one.
Read together, the columns say Parallel is easy to budget and easy to overspend on in one specific way, which is the default. Set the mode on every search and the processor on every task, and the bill is exactly what the price list says.
Who should actually buy this
Buy Parallel when you want agent search or web research priced per request, with the cost known before the call. Search in fast mode at $1 per 1,000 is among the cheapest agent search in this directory, and the free $5 a month covers 5,000 of those requests.
Use the Task API when a search is not enough and you need structured answers about accounts, and start on Lite or Base. Parallel describes Lite as basic retrieval of about 2 fields and Core as complex research of about 10, and the processors above Core as exploratory and deep research rather than enrichment.
Against Exa, the two are close on search and differ on what sits around it: Exa adds contact enrichment at $0.02 an email, Parallel adds Task and FindAll. Run the same queries through both before choosing.
Before you build against it
Five things to settle first, each one read off Parallel's own pages. The ticks are remembered in this browser, so you can come back to a half-finished integration.
Where the data comes from, and what you may do with it
The question that arrives late and expensively is who is allowed to hold the data. These are the fields it returns, the coverage it claims and the compliance it publishes.
| Provider | Fields returned | Coverage | Compliance | Redistribution |
|---|---|---|---|---|
| Parallel | Firmographics, Intent and news, Web pages | Global | SOC 2 | Not stated |
Parallel states no redistribution position on its pricing or docs pages. Internally that never matters. Put the rows in something your customers can see and it becomes the question your legal team asks first.
Two practical notes. A compliance badge describes the vendor rather than your use of it, so a GDPR DPA does not give you a lawful basis for the outreach you send. And the vendor runs its own opt-out, which means a person who asks you to delete their record has to be removed from your copy as well as from theirs.
Parallel against the obvious rival
Same rate cards, read side by side, with the arithmetic done on a worked volume rather than left to you: Exa vs Parallel. Every pairing in the directory is on the compare hub.
Calling these APIs one by one? Run them from one Clay table.
Clay runs a waterfall across 150+ data providers from one table, without a separate contract with each. A lookup that finds nothing costs no data credit, and the free plan includes 500 actions a month.
Frontal is a Clay Elite Studio Partner.
Nothing worth quoting yet
We searched Reddit for threads about the Parallel API and have found none worth quoting yet. When one turns up, it goes here, and nothing on this page leans on an unquoted paraphrase.
Parallel API, answered
Is the Parallel API free?
$5 of free credits every month, up to 5,000 requests, plus up to $80 at signup. Paid use starts at $1/1k searches.
How much does the Parallel API cost?
Parallel starts at $1/1k searches. Its unit rate is $0.005 per search, default mode. It bills pay as you go, with a free tier. Read from Parallel's own pricing page on 24 October 2026.
What are the Parallel API rate limits?
Search and Extract 600 a minute, Tasks 2,000, FindAll 300 an hour.
How do I get a Parallel API key?
Create a key in the Parallel platform and send it in the <span class="mono">x-api-key</span> header. Requests authenticate with API key in the x-api-key header. The current steps are in Parallel's own documentation, which is authoritative on anything that has changed since 24 October 2026.
Does Parallel have an MCP server?
Yes, at https://search.parallel.ai/mcp. It is listed with the others on the MCP page.