Tavily vs Exa vs Serper vs API Pick: Which Web Search API for LLMs?

Sarah Choy·2026年5月2日 發佈·約 9 分鐘閱讀

尚無翻譯，顯示英文版本。

Tavily, Exa, Serper, and API Pick all promise to be the search layer for your LLM. They differ sharply on output shape, filters, and how they bill. Here is a side-by-side from someone who ships agents.

一句話總結

• •Use Tavily when you want a single hosted RAG endpoint that returns LLM-ready answers — at the cost of opaque ranking and a subscription floor.
• •Use Exa for semantic-first discovery where neural ranking matters more than freshness; budget accordingly at scale.
• •Use Serper if you need raw Google SERPs and you will do your own cleaning, ranking, and snippet shaping.
• •Use API Pick Web Search when you want pre-shaped JSON snippets, transparent per-call credit pricing, country & date filters, and only pay for HTTP 200 responses.

What \"web search API for LLMs\" actually means

General search APIs like Google Custom Search, Bing Web Search, and SerpAPI return search-engine result pages — the same blue links and rich snippets a human would see. That format is wrong for a language model. An agent doesn't want to parse a SERP. It wants a small, ranked list of titles, URLs, and clean text snippets it can quote into a context window. The four APIs here all promise that, but they make different trade-offs in how they do it.

We will compare on five practical axes: output shape, filtering, pricing model, integration ergonomics, and what they don't do.

The contenders, in one paragraph each

Tavily

Hosted RAG-as-a-service. tavily.search returns ranked snippets; tavily.qna bundles search with a quick LLM answer. Strong fit for chat assistants where you want \"give the model an answer-ready blob\". Subscription-based with usage credits.

Exa (formerly Metaphor)

Neural / semantic-first index. Designed around \"find me URLs that look like this URL\" and embedding-based ranking, with options to retrieve highlights or full content. Strongest when freshness matters less than topical similarity. Subscription with credit overages.

Serper

Raw Google SERP API. Returns the JSON shape of a real Google search results page — organic, knowledge graph, places, videos. You do the snippet cleaning and ranking yourself. Cheap per query, but you ship the LLM-shaping layer.

API Pick Web Search

Pay-as-you-go semantic web search shaped for LLM tool calling. POST /api/search/web returns 5 (max 10) ranked results with titles, URLs, and pre-cleaned snippets, plus optional country_code and start_date/end_date filters. 15 credits per call (~$0.015), only deducted on success.

Side-by-side

Output shape: the part that matters most

The reason this category exists at all is that LLMs cannot reason effectively over a SERP HTML blob. They reason over short, named, ranked text. The single biggest predictor of whether a search API works well as an agent tool is therefore: how clean is the snippet?

Tavily and API Pick aggressively clean snippets. Exa returns either highlights or contents depending on flags — fine, but you decide how much to ask for. Serper hands you the raw SERP and assumes you'll run an extractor next. The latter is a reasonable choice if you already operate a content extractor; otherwise it is hidden work.

With API Pick, a typical response looks like:

{
  "results": [
    {
      "title": "Retrieval-augmented generation - Wikipedia",
      "url": "https://en.wikipedia.org/wiki/Retrieval-augmented_generation",
      "snippet": "Retrieval-augmented generation (RAG) is a technique that combines\nsearch with text generation, often using vector search to ground LLM\nanswers in retrieved documents."
    }
    /* …more */
  ],
  "result_count": 5,
  "credits_used": 15,
  "remaining_credits": 985
}

That shape drops directly into a function-calling response without further parsing.

Filtering: country and recency

Two filter dimensions matter for production agents:

• Country / locale: a financial agent in the UK should not get US-only sources by default.
• Date range: a market-research agent asking \"what happened this week\" must reject anything older than 7 days.

All four APIs expose some form of both, but expressivity varies. API Pick uses ISO date strings (start_date=\"2026-04-01\") which is unambiguous, vs. Google's coarser qdr buckets (past hour / day / week / month).

Pricing model: subscription vs. pay-as-you-go

Subscription-based APIs (Tavily, Exa) work well when you have predictable, steady traffic. They become awkward in three common patterns:

• You're prototyping and don't want a monthly commitment.
• Your traffic is bursty (e.g. a research agent that runs in batches).
• You build agents that retry aggressively on partial failures.

API Pick uses a credits model — $5 buys 5,000 credits; Web Search costs 15 credits per call; credits never expire and are only deducted on HTTP 200 responses. That last clause matters more than it sounds: an agent loop that retries five times on a transient 502 is free, not 5×.

Integration ergonomics

The lowest-friction integration is one where you can paste a JSON tool schema into your agent code without writing a wrapper. API Pick publishes ready-to-use schemas:

# OpenAI function tool schema
curl https://www.apipick.com/api/search/web/tool-schema

# Returns OpenAI tool definition + Claude tool use definition

With OpenAI Assistants:

from openai import OpenAI
import requests

client = OpenAI()
schema = requests.get("https://www.apipick.com/api/search/web/tool-schema").json()

assistant = client.beta.assistants.create(
    name="Research Agent",
    model="gpt-4o",
    tools=[{"type": "function", "function": schema["openai"]}],
)

With Claude tool use:

import anthropic
import requests

schema = requests.get("https://www.apipick.com/api/search/web/tool-schema").json()
client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=[schema["claude"]],
    messages=[{"role": "user", "content": "What's new in RAG research this week?"}],
)

What none of these APIs do

No web search API will reliably answer \"every X from this domain since 2019\". For deep-archive coverage you still want to pair search with a focused crawler or a domain-specific dataset. None of them dedupe across closely-similar URLs perfectly. And none of them solve the upstream problem of a stale or low-authority source — that's a content-quality decision your agent has to make.

Choosing fast