How to Choose an AI Web Search Tool
NanoGPT currently offers nine web search options. That is more choice than most people need for a quick lookup.
Most people can leave Linkup selected, use Standard Search, and move on. It is the default because it covers ordinary current-information questions without much setup.
Switch when your task has a clearer requirement. A search that finds breaking news may be the wrong tool for finding a paper, reading a pricing table, or collecting text from several pages.
First decide what you need back
“Web search” can mean several different things.
Search results give you titles, links, and snippets. This is enough for finding recent announcements, checking whether something exists, or giving a model a small set of sources.
Page content goes further by opening results and returning cleaned text. This helps when the answer depends on details that do not fit in a search snippet, such as documentation, product specifications, or a table buried halfway down a page.
Extraction asks for specific fields from a page: prices, names, dates, contact details, requirements, or another defined structure.
Research runs several searches, reads multiple sources, and produces a longer cited report. It is slower and costs more, but it can save work when the question is genuinely broad.
Choose the lightest tool that returns the material you need. Paying for scraped pages is wasteful when one snippet answers the question. A cheap search is not cheap anymore if it sends thousands of unnecessary words into the model afterward.
The nine options at a glance
These are NanoGPT prices as of July 18, 2026. They cover the search operation itself; the model that reads the results may add token costs. Options such as extra pages, more results, or research reports can change the final amount, so check the estimate shown in the app before running a large job.
| Tool | Typical NanoGPT search cost | Best starting point for |
|---|---|---|
| Linkup | $0.006 standard / $0.06 deep | A general default, sourced answers, structured results |
| Brave | $0.005 | Fresh, localized, language-specific web results |
| Tavily | $0.008 standard / $0.016 deep | News, finance, page content, images, balanced search |
| Exa | $0.005 instant/auto / $0.012 deep | Papers, GitHub, companies, semantic discovery |
| Kagi | $0.002 Enrich or news / $0.025 Full Search | Non-commercial enrichment, news, discussions, Kagi's full index |
| Perplexity | $0.005 | Controlled result and page-content budgets |
| Valyu | $0.015 standard / $0.075 deep at default result counts | Retrieval with explicit relevance and spend controls |
| Sofya | $0.01575 for search | Search, URL fetching, field extraction, research reports |
| Firecrawl | ~$0.0105 standard / ~$0.018 deep with scraping | Search plus cleaned page text, PDFs, research, GitHub |
Prices change. Treat this table as a comparison of the current shape of each option, not a promise about future billing.
Linkup: the default, and fine to leave alone
Linkup is the easiest place to start. Standard Search is inexpensive and handles ordinary questions that need fresh sources. It can return a result list, a sourced answer, or structured output on NanoGPT.
Deep Search is for questions where one round of results is unlikely to be enough. It also costs ten times as much as Standard Search, so do not leave it enabled out of habit.
Use Linkup when you do not have a provider-specific reason to choose something else.
Brave: conventional search with precise filters
Brave Search behaves like a conventional search engine with unusually precise controls. NanoGPT exposes country, search and interface language, freshness, Safe Search, spellcheck, result count, pagination, and extra snippets.
Use it for localized or recent results, such as French coverage from the past week or English-language reporting from the United States. Choose another tool when you need full page text, multi-step research, or semantic source categories.
Tavily: the balanced AI search option
Tavily sits between simple link search and a research pipeline. It supports Standard and Deep Search, general, news, and finance topics, date and country filters, images, an optional generated answer, and cleaned page content in markdown or plain text.
It fits current events, market updates, product comparisons, and questions where images or readable source text help. If snippets are enough, leave raw content off; extra page text increases the amount your model must read.
Exa: semantic discovery for specific source types
If your query describes an idea rather than a keyword, Exa is built for that. Its categories include research papers, companies, news, PDFs, GitHub, financial reports, people, tweets, and personal sites.
Instant and Auto start at $0.005, while Deep starts at $0.012. You can filter text and dates or request crawled context, but page contents add $0.001 per crawled page, so result count and subpages matter.
Kagi: three different indexes, three different jobs
Kagi exposes more than one kind of search in NanoGPT:
- Enrich for inexpensive non-commercial context enrichment
- News & Discussions for recent coverage and conversations
- Full Search for Kagi's complete search results
Enrich and News & Discussions currently cost $0.002 per request, while Full Search costs $0.025. That wide gap makes the source selection more important than the Standard or Deep label.
Use Kagi when one of those indexes matches the job. In particular, respect the non-commercial label on Enrich rather than treating the cheapest option as a universal default.
Perplexity: control how much content comes back
Perplexity Search lets you set the maximum number of results, a limit for each page, and a total text budget across the search. Country, language, and domain filters help keep the result set focused. This is useful in automated workflows where an unexpectedly large response could increase model costs or crowd out the rest of the task.
This is the Perplexity search provider, not the separate Perplexity Academic Researcher model. Use Academic Researcher when you want a literature-review workflow rather than a bounded set of web results.
Valyu: pay for the number of results
Valyu charges $0.0015 per result. Standard defaults to 10 results ($0.015) and Deep to 50 ($0.075), so Deep costs more by default than any fixed-price search in the table unless you lower the cap.
Use Valyu when you want spending to scale directly with the amount retrieved. You can control result count, price, relevance, response length, country, category, and whether to return only URLs.
Sofya: search, fetch, extract, or research
Sofya is closer to a small web-research toolkit than a single search endpoint.
On NanoGPT's Web Search page, it can:
- Search and return page content
- Fetch one or more known URLs as clean text
- Extract requested information from a page
- Run a multi-source research report
Choose Sofya when the job continues after discovery: search for relevant product pages, fetch the best candidates, then extract the same fields from each. In ordinary web-enabled chat, selecting Sofya uses Search. Fetch, Extract, and Research are separate operations; the Web Search page shows their estimates before you run them.
Firecrawl: search and scrape in one step
Sometimes the answer is inside the page, not the snippet. Firecrawl searches and scrapes in one request, returning up to ten pages as markdown with optional domain, GitHub, research, or PDF filters.
The default estimate is about $0.0105 for Standard and $0.018 for Deep with scraping enabled. Result count and scraping change the amount, and the returned text can increase the cost of the model that reads it.
Standard or deep?
“Deep” does not mean exactly the same thing for every search tool. It may increase search effort, result count, page reading, or some combination of them.
Start with Standard, Instant, or Auto when:
- The question has a specific factual answer
- You need one announcement, page, or recent event
- You are testing an automated workflow
- Cost and response time matter
Move to Deep when:
- The question spans several sources or viewpoints
- The first result set misses an important part of the topic
- You need stronger recall across a niche subject
- The cost of missing evidence is higher than the extra search cost
Do not use Deep to compensate for a vague query. Adding a date range, location, source type, or required domain often improves the result more cheaply.
For repeated or important work, run the same real query through two likely candidates. Compare source relevance, coverage, returned text, total model cost, and how much cleanup the answer needs. Search quality varies by topic, language, location, and date; a single benchmark cannot pick the right provider for every workload.
Search results are evidence, not proof
No search tool makes a model automatically accurate. A result can be outdated, duplicated, taken out of context, or drawn from a poor source. Extracted page content can also omit interactive elements, tables, or text rendered in unusual ways.
For consequential questions:
- Ask for links and publication dates.
- Prefer original documents over summaries about them.
- Check whether multiple results repeat the same underlying source.
- Ask the model to show where sources disagree.
- Open the most important source yourself before acting.
Web search also sends the query to a third-party search service. Do not put passwords, private documents, customer data, or other secrets into a search query. Fetching or scraping shares the URL and page contents too, so avoid private or internal links. NanoGPT's privacy page links the relevant provider policies.
Getting started
In chat, open Settings → Conversation → Web search provider, choose a provider, and enable Web for the conversation. The Web Search page exposes the advanced filters and shows the estimated search cost before you submit.
Start with Linkup Standard. Switch because the task calls for a different source type, output format, filter, or budget—not because a longer provider list makes the default look too ordinary.
