Perplexity has taught its Decisions API to look things up before it answers, closing the biggest gap in a fast, cheap classifier: facts that changed after training. Developers turn it on by adding “mode”: “web_grounding” to a request, according to Perplexity’s company blog.
The Decisions API returns probabilities instead of prose. A developer sends content as text, JSON, or images, along with one or more questions. A yes-or-no question comes back as the chance of yes, a multiple-choice question as a chance for each option, and a scoring question as a chance for each level of a rubric plus an expected score. That works for “does this review describe a defect?” It fails for “is this person still the CEO?”, because nothing in the input says whether the answer changed last month.
Grounding runs in Perplexity’s API gateway before the model sees anything. The gateway writes one Fast Search query per question, runs them in parallel, pools what comes back, removes repeated links, and attaches each page’s title, link, and body text to the input as evidence. If the input is itself a URL, that page is fetched as the evidence and the model reads it directly. Should a search fail or hang, the model simply uses whatever evidence did arrive. The response gains a search_results field listing the pages it read, so a developer can audit the basis for each number.
Perplexity reports two sets of results. In its own test, it put 328 claims, each either true or false, to one model in two runs. The claims were divided between October 2026 news and obscure Wikipedia facts, and each true statement was paired with a false copy altered in a single detail. Without grounding the API got 56.4 percent right, barely better than a coin flip. With grounding it got 97.3 percent, and errors fell from 143 to 9.
The cost is time. The typical reply slowed, with 139 ms becoming 366 ms, and the slowest five percent of replies climbed from 514 to 742 ms.
PWS-Decisions is a public benchmark whose 100 open questions cover sales-account screening, what products can do, rule checks, and breaking events. On it, Perplexity says accuracy went from 51 percent to 86 percent. It ran the grounded model (pplx-decider-v1.2) on 4 November and counted 86 exact-match answers. Perplexity also claims the highest public-set accuracy on the benchmark’s published leaderboard, and a typical response time below 400 ms.
Those claims come with a catch. Perplexity ran its own product on the benchmark, and the latency comparison mixes two measurement setups: 370 ms typical and 685 ms for the slowest five percent, measured from Perplexity’s own client over the 100 public questions, against 560 ms and 710 ms that the benchmark’s authors logged from their client across all 300. No independent run of the grounded model appears in the post.
Pricing is simple. Perplexity charges $0.001 for every finished search, on top of input tokens, and the retrieved evidence counts as input tokens too. A request can carry up to 100 questions, so the search fees for a maximal request come to about ten cents. Perplexity suggests three uses: checking whether accounts raised funding in the last 90 days, watching for a stable release newer than a pinned software version, and confirming an event is still on before an agent books it.
For a team scoring a 10,000-account list with one funding question each, search fees would run about $10 before token costs. That is cheap enough to try on a sample this week, and the 86 percent benchmark figure, not the 97.3 percent one, is the closer guide to how it will behave on messy real questions.
Reported by the Perplexity blog (perplexity.ai), which published the announcement without a listed date; benchmark runs cited are dated 29 October and 4 November 2026.
