Perplexity adds web search to its Decisions API
Perplexity is adding real-time web search to its Decisions API through the `web_grounding` mode. In tests, accuracy rose from 56.4% to 97.3% on current claims, although responses take longer and each search adds a cost of $0.001.
Perplexity has added real-time web search to its Decisions API, a tool that answers closed-ended questions with probabilities. With the new web_grounding mode, the model checks the internet before answering and uses those results as evidence.
This addresses a key limitation of AI models: their knowledge can be out of date. Information received during training is not enough to tell you, for example, whether someone is still a CEO, whether a company has just raised funding, or whether a product has a newer version.
How it works
The change for developers is small. You just add "mode": "web_grounding" to an API request.
Before calling the model, Perplexity's gateway creates one search for each question, runs the queries in parallel, and removes duplicate URLs. It then adds the titles, links, and text it finds to the context. The model responds in the same format as before, but with more recent information.
The content sent as context, called state, can be text, JSON, images, or a URL. In the last case, the model analyzes the specified page directly, so the developer does not have to copy its contents.
The response also includes search_results, allowing the application to review which sources the model saw. If a search fails or takes too long, the response is generated using the evidence that did arrive.
More accuracy, at the cost of some speed
Perplexity tested the feature with 328 true-or-false claims. They consisted of October 2026 news and uncommon Wikipedia facts, along with false versions that changed just one detail.
Without web search, the model got 56.4% of the claims right. With web_grounding, it reached 97.3% and reduced errors from 143 to 9.
The trade-off is higher latency, meaning more time before you receive the response:
- The median rose from 139 to 366 milliseconds.
- 95% of responses arrived in less than 742 milliseconds, compared with 514 milliseconds without search.
On the public PWS-Decisions benchmark, which contains 100 questions about current facts, accuracy rose from 51% to 86%. Requests with search took a median of 370 milliseconds in Perplexity's measurement.
What an application can do with this
The value lies in turning changing information into automated decisions. For example:
- Check whether a company has raised funding in the last 90 days.
- Detect whether a newer stable version of a tool is available.
- Confirm that an event is still active before booking it.
- Verify a policy or product capability before an agent takes action.
The result is not just a text response. The API returns a probability for each option, allowing the code to decide whether to continue, request human review, or stop the process.
Cost and limits
Each completed search costs an additional $0.001, separate from input tokens. The evidence found also counts toward the input limit, and a request can include up to 100 questions.
To enable it in Perplexity's Python or TypeScript libraries, pass the mode in the request body. In Python, for example, you can use extra_body={"mode": "web_grounding"}.
The important point is that a decisions API no longer has to choose between a fast but outdated model and a separate search managed by the developer. It can check the web just before answering, show the sources it used, and provide a probability ready for an application to act on. The next thing to watch is how it maintains quality when sources contradict each other or a page changes quickly.