Parallel for AI agents

Parallel is the provider an agent reaches for when a question is too deep for a search. It has an ordinary search endpoint and an ordinary extractor, and both are useful as a second source, but the piece that has no substitute in the network is the task: hand it a research question, come back later, and receive a written answer with citations. It settles over x402, which is why an agent can call it with no account anywhere.

What Parallel does well

Deep research as one call. parallel/task takes a question and a processor tier, plans the search, reads the sources, and returns a synthesized, cited result. It is asynchronous: the call returns a run id, and the agent polls until it is complete. For a multi-hop question this replaces a loop the agent would otherwise have to orchestrate itself.

A second index. parallel/search is a different engine from Exa. In a corroboration pipeline, searching both and deduplicating catches what either misses.

Objective-guided extraction. parallel/extract takes an objective alongside the URLs, so the extraction focuses on, for example, pricing tables rather than returning the whole page.

No account, no key, no card. Parallel accepts x402 payment per request. Vaaya pays it that way on the agent's behalf, so the agent sees a price and a ceiling and nothing else.

Actions and prices through Vaaya

Priced before the call, billed only on success, under max_cost_cents.

Example request

{
  "service": "parallel",
  "action": "task",
  "params": {
    "input": "Compare Stripe's Machine Payments Protocol and x402 on settlement, buyer requirements, fees, and status. Cite primary sources.",
    "processor": "pro"
  },
  "max_cost_cents": 10
}

Returns a run id. The agent polls parallel/task-status at no cost until the status is complete and reads the answer. Setting processor: "ultra" with a ten-cent ceiling is refused before anything runs, with the actual price returned.

When to use Parallel

Pick the task when the question is genuinely multi-hop and a cited answer is the output. Pick the search as a second source next to Exa when a result needs corroboration. Pick the extractor when you want extraction focused by an objective.

Do not pick the task for one-hop questions; a chain of one-cent Exa searches and reads usually reaches the same answer for less. Do not pick Parallel's extractor for large batches of URLs where you only need the text; Exa's contents endpoint is a tenth of a cent per URL. For a written answer to a simpler question, Linkup's sourced answer is a cent.

Usage notes

The task is asynchronous; do not poll in a tight loop, even though status checks are free, because a task can take minutes. Reserve ultra for questions that have failed at pro. Cap every task with max_cost_cents set to the tier's price so a wrong processor value cannot cost more than intended.

Parallel's documentation is at parallel.ai. The deep research recipe is at /recipes/search-scrape.

Questions

What is the Parallel API?

A web research API for agents with three parts: a search endpoint, a per-URL content extractor, and an asynchronous task that runs multi-step research and returns a cited result. Through Vaaya all three are reachable with one key and priced per call.

How much does a Parallel research task cost?

Ten cents at the pro tier and thirty cents at ultra, per task, through Vaaya. Polling the task's status is free. Search is a cent a call and extraction is a cent per URL.

When should an agent use Parallel's task instead of chaining searches?

When the question needs several hops that would take an agent ten or more cent-level calls to assemble, and a single cited answer is the deliverable. For one-hop questions, a chain of Exa searches and reads is cheaper.