Jev says no, the web says yes: Parallel forces decision models to check today's data

Decision models have always had an awkward problem: they answer quickly, but they don't know what happened yesterday. Parallel's newly launched Decisions API aims to solve this gap by squeezing real-time web retrieval and structured judgment into a single call. Official pricing is $0.001 per request, with 5,000 free credits included.

Parallel Decisions API

Decision models are suited to small outputs: yes or no, choose one from a list, classification, routing. The problem is that answer quality depends on context. If an answer depends on recent changes or information that exists only on the web, a normal pipeline has to be split into retrieval, context fetching, and then letting the model make a judgment. Parallel combines the two steps: you provide state and named questions in the request, and the API searches the web for evidence, returning typed answers with probability, confidence, and sources. The default mode is fast; you can ask up to 100 related questions at once, sharing the same set of retrieval evidence, and each question including the shared state can contain up to 32,000 estimated input tokens.

A telling comparison is European data residency. Parallel announced support on October 14. When Jev was asked, it answered no, with probability 0.86, confidence 0.72, and no sources. Going through the Decisions API, the answer was yes, probability 1.00, confidence 1.00, citing parallel.ai's blog and docs. It's not hard for a model to answer a multiple-choice question; the hard part is knowing that yesterday's web pages may no longer count.

Batch scenarios are even more direct. Given one shared context and four questions for Vercel, the API gathered evidence from 10 sources, returned in 681 milliseconds, and cost $0.001. The employee-count bracket was 501_plus, with confidence 0.55, and it was routed to human review; whether it has a public API, whether it has an enterprise edition, and whether it is an agent product all had confidence 1.00. Low confidence does not force a label, which is more useful than returning one more pretty boolean.

Parallel itself used it for two internal projects. In a financial KYB scenario, 426 public companies were rechecked with the Decisions API in under ten seconds, costing 85 cents, and 98% of the flags held up. In GTM lead routing, 500 inbound leads were processed in under nine seconds, costing about 50 cents, and 97% of self-serve routing landed in the correct segment. What really makes these two numbers useful is that the per-item cost is low enough to run at batch scale.

Usage is not complicated: get an API key and send state and questions to POST /v1/decisions; the noul type returns the probability that the condition holds. Each question may return a refusal, which is a successful response, not an HTTP error; check type before reading the answer. citations belong to the entire call and are not mapped one by one to specific answers, so you have to design the audit trail yourself.

If you don't want to open an account first, you can also use Parallel MCP: connect the client to https://mcp.parallel.ai/mcp and call web_decisions, with no account or API key required. It's suitable for validating a use case first, then deciding whether to put the Decisions API into a production pipeline.

My judgment is that this kind of API will productize the dirty work of search plus decision-making, but it will not make the retrieval pipeline disappear. Probability and citations make decisions auditable and give engineering teams another set of metrics to watch. Confidence scores now have to be tested in the open web, especially when multiple questions share the same set of retrieval evidence. If your workflow has lots of small judgments and the answers change with the web, $0.001 per call is worth trying; if you need answer-level citations and strong explanations, you'll still need to add a layer yourself.

Links: