Parallel Web Systems' Decisions API Answers 100 Questions for $0.001

Parallel Web Systems has launched the Decisions API, a classification service that retrieves fresh web evidence before returning typed answers. A developer sends context plus one or more named questions, and the API responds with probabilities and source citations. The design targets decisions that depend on facts newer than a model’s training data.

Calls cost $1 per 1,000 requests, or $0.001 each, with 5,000 free requests in the starting allowance. One request can contain up to 100 related questions that share retrieved evidence. Parallel estimates that this batching costs roughly one-tenth as much as combining a standalone decision model with a separate retrieval pipeline. Actual savings will depend on batch size, retrieval needs, and the existing stack.

Live evidence enters the loop

Fast classification models such as Parallel’s Jev can answer in a few hundred milliseconds, making them suitable for decisions inside latency-sensitive agent workflows. Parallel calls these “System One” models, borrowing a term for quick judgments. Their answers remain limited to training data and the context supplied at runtime.

Parallel demonstrated that limitation with a question about whether the company supports European data residency. The feature had recently shipped, but Jev answered no with 0.86 probability because its training data predated the announcement. The Decisions API retrieved the company’s current blog and documentation pages, then answered yes with probability 1.0.

Each API call includes source links that make the retrieved evidence inspectable. Attribution remains call-level: the top-level citations array covers the entire request and does not connect individual sources to specific answers. A score of 1.0 records model confidence, so applications still need thresholds and review policies appropriate to the consequences of an error.

One call, three answer shapes

A request contains a shared state and a questions dictionary. Each named question specifies an output type and instructions.

The quickstart documents three question types, which can be mixed within one request.

curl -X POST https://api.parallel.ai/v1/decisions \
  -H "x-api-key: $PARALLEL_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "state": {
      "company": "Vercel",
      "domain": "vercel.com"
    },
    "questions": {
      "has_public_api": {
        "type": "noul",
        "instructions": "Does the company offer a public API?"
      },
      "employee_band": {
        "type": "noul",
        "instructions": "Is headcount above 500?"
      }
    }
  }'

Responses require type-aware handling because a question can produce an answer or a refusal object. A refusal arrives within a successful HTTP response, so checking the status code alone is insufficient. Client code should inspect each result’s type before reading its value or probability.

• Answers: Completed classifications include a probability.
• Citations: Sources appear once in a top-level array for the complete call.
• Refusals: Applications must detect these in the response body and decide whether to retry, escalate, or stop.

Batching cuts the unit cost

At list price, a request costs $0.001 whether it contains one question or the documented maximum of 100. A full batch therefore has a nominal API cost of $0.00001 per question. The practical gain depends on whether the questions concern the same entity and can use the same evidence.

Parallel’s Vercel demo bundled four checks covering employee count, public API availability, enterprise plans, and agent products. The request retrieved 10 sources, returned four probability-bearing answers in 681 milliseconds, and cost $0.001.

Parallel also reports two internal benchmark runs. The published results provide scale, total time, cost, and outcome rates.

The reported costs equal about $0.002 per company in the KYB run and $0.001 per lead in the routing run. The figures are vendor-reported, and the published summary omits the datasets, ground-truth process, abstention rate, concurrency settings, and comparison baselines needed for independent reproduction. The elapsed times describe complete batch runs; per-record latency depends on the concurrency configuration.

A checkpoint inside agent workflows

Parallel has spent about a year building web-grounded services including Search, Extract, Task, Monitor, and the Responses API. Decisions adds typed classification over retrieved web evidence. The company also raised a Series B at a $2 billion valuation to fund its agent-infrastructure products.

Common integrations place the API before a costly model call, a database update, or an external action. The application can proceed above a confidence threshold, retry with revised context, or send uncertain results to human review.

Parallel also exposes the service through its Model Context Protocol server as the web_decisions tool. MCP gives compatible agent clients a standard way to invoke external tools. Parallel says Claude Code and other MCP-aware clients can use this tool without an API key at lower usage limits, while production applications can call the authenticated HTTP endpoint directly.