Give an agent a search box and you've made a quiet architectural decision you probably didn't notice. "Search the web" sounds like one capability, but the APIs that provide it deliver the web at wildly different stages of doneness — and the stage you pick determines how much code you write, how many tokens you burn, and what you pay per query. The vendors don't really compete on finding better pages. They compete on how much of the reading they do before they hand it back.

Line them up and the market sorts into four rungs. It is the same web underneath; the difference is how far down the crawl-clean-think pipeline the API delivers — and, almost perfectly, the price climbs each rung as more of that work moves from your process to theirs.

The bottom rung is the classic SERP API — Serper wrapping Google's results, or the Brave Search API serving its own independent index. You send a query, you get back links and snippets: titles, URLs, a sentence of context. That's it. Whatever the agent actually needs to read lives behind those links, which means you still have to fetch each page, strip the nav and cookie banners, and feed the cleaned text to your model yourself.

This is the cheapest rung by an order of magnitude — raw SERP queries run a small fraction of a cent each — precisely because it does the least. You are renting a result list and bringing your own crawler, your own cleaner, and your own reasoning. For an agent that already has a page-reading layer and just needs to know which URLs to point it at, that division of labor is exactly right. For one that doesn't, the cheap query is a trap: you've bought the easy half and still owe the hard one.

Rung two: cleaned content, ready to read

Tavily was built for the next rung up. It is a search API designed for LLMs and RAG: you send a query and get back not just links but the content — extracted, cleaned, and shaped for a model to read — with a search_depth knob (basic or advanced) that trades cost for thoroughness and an optional LLM-generated answer if you want one. There's a separate extract endpoint for pulling clean text from URLs you already have. The pitch is that crawling and cleaning are undifferentiated work, so the provider should absorb them and hand you model-ready text.

That this rung is now considered strategic infrastructure isn't speculation. In February 2026, Nebius — a public AI-cloud company — announced it was acquiring Tavily to add "agentic search" to its platform, a deal reported around $275M. When a cloud provider buys a search API the way it would buy a database, the signal is clear: feeding the web to agents has graduated from a developer convenience to a layer of the stack worth owning.

Rung three: search by meaning, not by keyword

Exa (the former Metaphor) climbs sideways rather than up. Its bet is that keyword search is the wrong primitive for an agent in the first place: Google matches strings and reranks, while Exa runs a neural index that matches pages by meaning — "search the way an LLM would think about it." Ask for "companies using reinforcement learning for drug discovery" and a neural engine can return the right pages even when none of them contain those exact words, a query shape that makes a keyword API flail. It pairs /search with a /contents endpoint to clean the matches and an /answer endpoint if you want synthesis, plus "Websets" for structured bulk results.

The honest caveat is that neural isn't strictly better — it's better at conceptual recall and worse when you know the exact term and want an exact-match lookup, which is why Exa exposes a keyword mode too. And notably, Exa has stayed independent while Tavily was absorbed, a small tell that there's more than one viable theory of what this layer is. The point of the rung is recall by meaning; you still bring the reasoning, or pay a little more for /answer to do it.

Rung four: just give me the answer

The top rung skips the middle entirely. Linkup and Perplexity's Sonar API don't return pages for your model to read — they return a finished, cited answer. The provider runs the search, reads the results with its own model, and hands back prose with sources attached. Linkup leans hard on factual grounding and claims a state-of-the-art 91% on the SimpleQA benchmark — though that's a vendor self-report, the kind of number to verify against the eval before you build on it, not take as settled.

This rung is the most expensive per call because the provider is now paying for the crawl, the cleaning, and the answer-synthesis LLM — the entire pipeline as a single billed unit. The trade is real: you write almost no retrieval code and you also surrender control over how the answer was assembled and which sources it weighed. For a quick factual lookup mid-conversation it's the least work in the space. For anything where you need to inspect or re-rank the evidence, it's a black box you're paying a premium to not see into.

Nobody here is selling you a better web. They're selling you the web at four different stages of doneness, and the price is just a meter on how much reading they did before you got it.

Buy the rung, not the brand

The four products feel like competitors and mostly aren't — they're four answers to how much of the pipeline you want to own. So decide that first:

The expensive mistake is choosing by demo. The "give me the answer" rung dazzles in a sandbox and then hides the sourcing you needed in production; the cheap SERP rung looks like a bargain until you've rebuilt a crawler and a cleaner to make it usable. Price tracks doneness here more honestly than almost anywhere in the agent stack. Figure out how much of the reading you actually want to do — and buy exactly that much.