Introducing the Parallel Decisions API

gist

Parallel launched a Decisions API that answers yes/no and multiple-choice questions from the live web, with citations, at $1 per 1,000 calls. A decision model is a fast classifier that returns a typed answer with a probability. Parallel says models like Jev know only their training data and the prompt; in its own example, Jev said Parallel had no European data residency, while the API answered yes and cited two pages. One request can batch several questions that share context for the price of one call. All benchmark figures in the post are vendor-reported.

ideas

• Web grounding fixes stale answers from decision models. Parallel says System One decision models such as Jev cannot answer questions that depend on recent facts, like where mortgage rates closed, because they see only training data and the input.
• Every answer carries a probability and citations. The API returns a typed value, a probability or confidence score, latency, and source URLs, which Parallel presents as a way to audit answers from models that do not justify themselves or say "I don't know".
• Batched questions share one retrieval. In Parallel's Vercel example, four CRM enrichment questions under one shared company context came back in 681 ms for $0.001 from 10 sources, with the low-confidence employee band (0.55) routed to human review.
• Parallel reports pilot results on its own data. By Parallel's account, a know-your-business recheck of 426 public companies ran in under ten seconds for 85 cents with 98% of flags holding up, and 500 inbound leads were routed in under nine seconds for about 50 cents with 97% landing in the right segment.
• The API is also an MCP tool. The Parallel MCP server exposes it as a tool named web_decisions, free to use without an account or API key, with higher limits for signed-in users.

quotes