What happened
Parallel announced PWS-Decisions on November 1, 2026, describing it as a benchmark for web-grounded decisions. The set contains 300 closed-ended questions modeled on decisions developers may route through the company’s Decisions API: qualifying accounts, checking product capabilities, confirming policy rules, and tracking fast-moving events. The questions are designed to depend on facts that are current, niche, or scattered across the web. Parallel says each answer was confirmed through independent web research and a direct source, and that each system is scored using exact match under the same rules.
The company reports that Parallel Decisions API scored 78% on PWS-Decisions, compared with 41.7% for OpenAI Decisions and 33.4% for Jev. The post contrasts those results with JevBench, a benchmark of 231 classic decision tasks answerable from the prompt: there Parallel Decisions scored 89.6%, OpenAI Decisions 90.0%, and Jev 88.3%. Parallel argues that this difference matters because a model that answers from information already supplied in a prompt is not necessarily capable of producing a correct decision when it must obtain current facts from the web.
The benchmark is available for others to run on their own models. Parallel links both the code on GitHub and the dataset on Hugging Face. It also says users can try the Decisions API in its playground with 5,000 free requests. The API returns a typed, cited answer for a question such as yes/no, a choice from a defined list, or a score. Parallel positions it for decisions that need current information, including vendor validation, lead qualification, fact-checking, content moderation, and entity matching.
The release also makes a cost argument: an evaluation of decision systems should account for search as well as the model, because a low-cost model combined with a separate search call can cost more than a single Decisions API request. The post says requests can include up to 100 questions and cost $1 per 1,000 requests. Those details describe Parallel’s service, not a general price comparison for every possible system. The company’s benchmark post does not publish a detailed independent audit or establish that its API is the best choice for all decision workloads.
What the archive adds
This is a new fact on a known story, not a new product launch. The archive covered Parallel Decisions API: web-grounded classification, including its live-web grounding and pricing. PWS-Decisions adds a public evaluation set intended to let teams test the API and competing systems on current-web questions.
Trend signal
This supports agent economics moves to completed work only in a limited sense: Parallel argues that cost should include the search needed to reach a usable decision, rather than counting model calls alone. The benchmark itself measures exact-match accuracy, not verified task completion or end-to-end workflow economics, so it does not establish the pattern’s broader claim.
Why now
The benchmark follows Parallel’s Decisions API launch, which the company’s post describes as having taken place a couple of weeks earlier. The release gives prospective users a way to inspect the API’s claimed advantage on live-web decisions and run the same questions against their own model. The source does not explain a broader market or technical event that prompted the timing.
Connections to other events
Parallel’s September Search Capability Leaderboard compared models on web-search tasks, while the October Decisions API announcement introduced a product for structured answers grounded in live web data. PWS-Decisions moves the comparison from general search capability to closed-ended decisions that depend on current sources. These are related evaluation and product developments, but the new benchmark does not validate the leaderboard or establish that search ranking alone predicts decision accuracy.
Why it matters
Teams often evaluate a model on questions with evidence already present in the prompt, then deploy it in a workflow where current facts must be found, checked, and converted into a bounded output. PWS-Decisions makes that mismatch visible as an evaluation problem. A team deciding whether to automate vendor checks, lead triage, or another repeated classification can use the dataset as a starting point, inspect whether its questions resemble its own workload, and compare systems under common scoring rules.
The results should not be read as a universal ranking: they are reported by the vendor whose API leads the test, and 300 questions cannot represent every domain, language, or source environment. Still, public code and data make the claim more testable than a closed product demo. The practical value is not simply that one API scored higher; it is the reminder to evaluate the full retrieval-to-decision path, with current evidence and the actual output format, before putting a model into an operational pipeline.
Limits and uncertainty
The source does not provide an independent reproduction of the scores, confidence intervals, detailed per-question breakdowns, or enough information here to assess how representative the 300 questions are of production traffic. It reports exact-match accuracy, which does not by itself measure citation quality, source freshness, latency, error severity, or the cost of human review. The company’s cost argument is framed around its own API and a particular type of request; the post does not establish comparable total costs for every competing architecture. Results may also change as web content and evaluated systems change.
Skill to build now
Build a small evaluation set from one repeated workflow this week: collect representative questions, define an allowed answer schema, and record the current authoritative sources a reviewer would accept. Test whether each answer depends on live information or can be answered from a fixed prompt. Then compare your current approach with a web-grounded system using the same questions and scoring rules, while separately tracking source correctness, latency, and cost. PWS-Decisions can provide a template, but your own cases should determine the deployment decision.
Reader takeaway
A benchmark of prompt-answerable questions is not enough when a production decision depends on current facts outside the prompt. PWS-Decisions offers an open set for testing that distinction, but its vendor-reported results are a starting claim, not independent proof. Evaluate your own live-web decisions, evidence quality, and all-in cost before choosing a system.