Questions & answers

Google's Custom Search JSON API is closed to new customers, so new projects need another option. SerpApi returns Google results as data. Its free plan covers 250 searches a month, and its Starter plan is $25 a month for 1,000 searches. Check its terms before you store or display the results, and confirm the plan that fits your volume before you build on it.

Many providers allow commercial use, but the terms decide it. Each provider sets rules for storing results, caching them, showing links and reselling output. Microsoft's Grounding with Bing terms, for example, require website URLs and search query links to be shown in the form Microsoft provides. Read the terms for the plan you buy, and get a legal review if your product stores or republishes results at scale.

Rate limits differ by plan. Brave Search API lists 50 queries per second on its Search plan and 2 per second on its Answers plan. Parallel's limits are per minute, with 600 per minute on its Developer tier and 1,500 on Business. Agents can fire many searches at once, so set a queue, cache repeated queries and read the limit before you pick a plan.

Grounding means that each claim in a model's answer traces back to a source the model retrieved. Good grounding returns the passages that support a claim and keeps the link to the page. Microsoft's Grounding with Bing Search is one example, but it does not return the raw results to developers. You can test grounding by checking each cited sentence against its page.

BrowseComp is a benchmark of questions that are hard to answer without browsing several sites. Vendors publish scores because the test rewards finding and combining facts from the web. Scores depend on the model, the step budget, the grader and the search tool, so one number is not enough. Hyperresearch's October 2026 table puts Hyperresearch Answers at 89% and Parallel at 73%, both vendor-run figures.

Yes. Perplexity's Search API bills $5 per 1,000 successful requests, and its Fast Search option is $1 per 1,000. Each successful request counts as one billing unit, even if it contains up to five queries. Its Orchestrator API bills tokens at each model's rate. Check the current price table before you budget, because model rates change.

It is a tool inside OpenAI's Responses API. You add a web_search entry to the tools list, and the model decides whether to search. Search actions are billed as tool calls on top of model tokens, so one answer can cost more than the search alone. OpenAI's pricing page lists current rates. Hyperresearch's October 2026 table lists the tool at 58% on BrowseComp, a vendor-run figure.

An MCP search server exposes search as a tool that AI apps can call through the Model Context Protocol. You.com lists a web search endpoint through its MCP server, with 100 free queries a day and no API key. Hyperresearch Answers is also available as an MCP server. Check which tools the server exposes, what each call costs, and whether results come with quotes or only links.

Jina gives each new API key a grant of 10M free tokens. Its Reader endpoint works without a key for basic use, while its Search endpoint asks for a key and charges a fixed token minimum per request, starting at 10,000 tokens. Jina prices in tokens, so check its current price table before you budget, and read the terms if you plan commercial use.

Some do and some do not. A provider may crawl its own index, license one from another company, or query a search engine and rank the results itself. The choice changes coverage and freshness. Ask which index a provider uses, how often it updates, and whether it covers your topic, such as filings, government data, academic papers or local news.

Some can. Hyperresearch Answers covers the open web plus 90+ structured sources, including SEC EDGAR filings, and returns answers with quotes from those sources. Many general search APIs return web pages instead, so check whether a provider covers the datasets your work needs before you build around them.

Build a fixed set of real queries from your product, and write down the answer you expect for each. Run every candidate API against the same set, with the same model and prompt. Score whether each answer is right, whether each citation supports its claim, and what one completed answer costs. A short pilot on your own queries tells you more than a vendor benchmark.