Questions & answers

It depends on the tier you call. Linkup lists $0.005 to $0.006 per search request. Perplexity lists $5 per 1,000 Search API requests and $1 per 1,000 for Fast Search. Parallel lists a starting rate of $1 per 1,000 requests. Exa lists $4 per 1,000 lite searches. Compare the tier that gives your app the quality it needs, not the lowest headline number.

For many questions, yes. Excerpts keep the context small and cheap to read. Full page text is better when the claim sits inside a table, a footnote or a long policy. A common pattern is to start with excerpts, then fetch full text only for the sources the agent actually cites. Test on your own questions, since the right mix depends on the topic.

Fewer than you might expect. Ten results is a common default, and Perplexity and Tavily cap a request at 20 results. More results add cost and tokens, and they can add noise. Start at five to ten, read the top few, and raise the count only if your pilot shows that relevant sources are missing.

Some do, in different forms. OpenAI's web search returns url_citation annotations inside the answer. Exa's answer endpoint and Hyperresearch Answers return the source URL for each quote. Hyperresearch checks each quote word for word against its page and marks it verified. Raw result lists give you links only, so your app must do any citing itself.

Build a set of 50 to 100 real questions from your users, record the source or answer you expect for each, and run every API on the same set. Score whether the right source appeared and whether the final answer is supported by it. Repeat the test every month, since indexes and rankings move. Public benchmarks help you shortlist, but they do not replace a test on your own questions.

Search fees are only part of the cost. Your model reads whatever the search returns, and that reading is billed in tokens. Brave's LLM Context lets you set a budget from about 1,000 to 33,000 tokens. Perplexity caps total and per-page content. OpenAI bills search content tokens at each model's input rate. Set a token budget before you ship, or your costs will track page length.

Often, yes. Exa filters by publish date with start and end parameters. Tavily accepts a time range such as day, week or month, or explicit start and end dates. Brave offers a recent-content filter. You.com returns a page age you can sort or filter on. We did not confirm a date filter in Perplexity's quickstart, so check its API reference.

Treat search results as untrusted input. A web page can contain text written to steer a model, so an agent should not gain new powers because a page asked for them. Keep any tool that sends messages, spends money or changes data behind an explicit approval step. Log the sources the agent used, and review them when an answer matters.

Hyperresearch Answers is our editorial pick for agents that need short answers with checked quotes. In its own October 2026 runs, it scored 89 on BrowseComp against 73 for Parallel Pro. Parallel is the runner-up for excerpts and speed modes. Exa suits neural search over links, and Brave suits teams that want an independent index. Pilot two options on your own questions before you commit.

Microsoft retired the Bing Search APIs on August 11, 2025. Its announcement said the product could no longer be used and was closed to new customers, and it pointed developers to Grounding with Bing Search in Azure AI Agents. Old code that calls it will not work. Plan a move to a current search API, and test result quality on your own questions before you switch.

Yes. Answers runs as a remote MCP server at answers.hyperresearch.ai/mcp. Its ask tool returns one short answer with quotes, and its search tool returns ranked sources with passages. If no source supports an answer, it reports not found. New accounts get free credit to start. Add the server to your client, then check the first few answers against the cited pages.

Check who ran the test, how many questions it used, and which judge graded the answers. A score from one vendor's own run is a claim, not a neutral result. Look for independent evaluations that run the same agent on each provider. Small samples can swing a lot, so treat gaps of a few points with care, and test on your own questions.