Perplexity adds live web grounding to its Decisions API

Perplexity has added optional live web searches to its Decisions API, which returns structured probabilities for decisions rather than prose. The change gives developers a way to supply fresh evidence to automated decisions, with measurable gains in a company-reported test and a few practical costs to keep in view.

Perplexity Watch analysis

What happened

The feature, available immediately according to Unite.AI, uses Perplexity’s Fast Search before the model answers. Developers can set a URL to guide the search, inspect returned sources in a searchresults field, and provide up to 100 questions in a grounded request. Searches can run concurrently, with a one-second limit per search and a three-second overall limit.

Perplexity tested the grounded model on 328 true-or-false claims drawn from October news and rare Wikipedia facts, with no evidence supplied in the initial prompt. The company reports accuracy of 97.3% with web grounding, against 56.4% without it. Median latency rose from 139 milliseconds to 366 milliseconds. These are results from Perplexity’s test, not a general measure of performance across developers’ own workloads.

Why it matters

The API’s structured answers can be used as thresholds in classification, routing and other automated decisions. Live evidence could help when a decision depends on something that changes, such as a software release or whether an event is still scheduled. The returned sources also give developers something inspectable alongside the probability, rather than a number with its workings tucked out of sight.

Grounding is not a free accuracy button. Searches cost $0.001 each, and retrieved text adds input-token charges and counts towards the 256K input limit. Search failures may leave a response standing: if no search succeeds, the model can answer from the original state and return an empty searchresults array. That distinction matters when a downstream system treats a probability as a green light.

Our read

This is a useful addition for developers whose decisions need current information, and the reported accuracy jump is striking. The right reading is not “the API is now 97.3% accurate”, but that Perplexity saw that result on a specific test, with higher latency and a search process whose success developers need to check. If you build on it, inspect the evidence returned and decide what your system should do when it is missing.

What to watch

• Whether independent tests reproduce the accuracy gains on real-world decision tasks.
• How often searches time out or return no results in production.
• Whether the Python and TypeScript SDKs add typed support for mode and searchresults.
• How search and token costs scale when questions are grouped or evaluated separately.

Discussion spark: For automated decisions, would you accept slower responses and extra search costs for fresher evidence, or should a system stop whenever that evidence is missing?

not affiliated with or endorsed by Perplexity