Search for Agents Is Becoming Its Own Market

TL;DR

• The September 11–13 measurement set spans ranked retrieval and managed agent work across Exa, Perplexity, Parallel, and Amazon Bedrock AgentCore. No provider leads every layer.
• Search, extraction, generated answers, research agents, task groups, monitoring, embeddings, and MCP integrations sit behind different endpoints and billing units even when vendors market them under one search label.
• Tavily, a search-for-agents startup that had raised $25M total, was acquired by Nebius for $275M in February 2026.
• No analyst firm tracks “search API for agents” as its own category yet. The closest proxy, Fact.MR’s agentic tool-use/API-integration estimate, puts the whole adjacent category at $6.9B in 2026 — and that’s every agent-tool-integration dollar, not search specifically.
• The pricing split is the market signal: raw retrieval is billed per query, while managed work is billed per successful run, agent effort, model tokens, tool calls, checks, or matched entities.

The September 11–12 retrieval measurements cover four search APIs built for AI agents — Exa, Perplexity’s Search API, Parallel, and Amazon Bedrock AgentCore Web Search — across 30 research queries, live filters, page-fetch behavior, and a fixed-model review of 600 saved results. No provider won outright:

Exa returned the most primary-source evidence. Perplexity returned the most useful results at the lowest search cost. AgentCore was fastest by a wide margin in this sequential run; the run does not establish why.

The complete measurement set also covers generated answers, research agents, stateful task runs, entity discovery, monitoring, embeddings, and MCP integrations. Those surfaces do not map cleanly to the “$X per 1,000 queries” number on a search pricing page.

An independent benchmark measures the task, not the query

The Artificial Analysis Search Index, with latest data from September 8, benchmarks 20 search products across 10 providers. Its methodology holds the candidate model, grader, prompt budget, and agent loop constant: GPT-5.6 Luna at medium reasoning, up to 25 turns, web_search and web_fetch, and a maximum of ten results per search. The index equally weights DeepSearchQA, BrowseComp, and AA-Omniscience.

Perplexity medium leads this task-level benchmark, while other configurations occupy different quality, cost, and latency positions. The model-only baseline makes the larger result clear: every listed search configuration lifts the same agent substantially above its internal knowledge.

This is a different measurement from the 30-query retrieval table above. Artificial Analysis scores the answer produced by an agent that can search and fetch repeatedly, and its cost columns include both search and model work per benchmark task. The retrieval table scores the evidence returned by one configured search request. AgentCore Web Search is also absent from Artificial Analysis’s documented provider settings, so the external benchmark independently covers three of the four providers in my retrieval comparison rather than replacing it.

The two views support the same selection rule: raw query price does not predict completed-task quality or total task cost. Better evidence can reduce model work; a cheaper search call can force more turns, tokens, or fetches.

The money says this is a real market, not a feature

Line up the last two years of funding across the vendors in this space and the pattern is infrastructure-company money, not feature money:

Exa’s Series C came eight months after an $85M Series B. Parallel — founded by Parag Agrawal — raised two $100M rounds five months apart. Perplexity’s valuation ladder went from $520M in January 2024 to $22.6B in January 2026 before the reported Nvidia talks, on the back of annualized revenue that tripled from under $250M to over $750M in about seven months.

The clearest signal is Tavily. It was a two-year-old company that had raised only $25M total when Nebius acquired it for $275M in February 2026. Tavily’s founder framed the deal in exactly these terms: “agentic search is a multi-billion-dollar opportunity, and we believe the market is poised to grow exponentially as enterprises deploy autonomous AI systems.” An AI-infrastructure company didn’t buy a feature. It bought a search layer to bundle into its own cloud platform.

None of this shows up if you only compare list prices. It shows up when you ask who’s writing the checks.

Nobody has drawn the category boundary yet

There is no established dollar figure for “search API for agents” as its own line item. The closest analyst estimate, Fact.MR’s agentic AI tool-use and API-integration market, sizes the broader category — every agent tool call, not search specifically — at $6.9B in 2026, growing to $39.6B by 2036.

That gap matters more than it looks. When a category has no agreed boundary, every vendor gets to draw its own — which is exactly what’s happening in the next section.

It’s four different businesses, not one

Line up Exa, Tavily, Perplexity Search, Parallel, SerpAPI, Bing Grounding, and Google’s Search grounding side by side and they stop looking like competitors. They’re four different products that all get called “search API”:

Tavily’s own docs describe the first row precisely: “the web layer for AI agents.” It also has a named partnership as LangChain’s recommended search tool — the strongest adoption signal I found for any single vendor. The second row is easy to mistake for the first until you read a direct comparison: “A SERP API returns the raw results-page JSON… An agent search API… returns a short list of already-cleaned, ranked context… They solve different problems and aren’t interchangeable.” SerpAPI isn’t competing with Exa. It’s a different tool that happens to share a category label. And Glean has nothing to do with the open web at all — it sits in the same sentence as Exa in a lot of vendor comparisons and shouldn’t.

If you’re picking a vendor by scanning a feature matrix, you’re comparing across all four rows at once without realizing it.

The pricing model is splitting in two

Three of the four vendors I tested sell raw search beside an agent or task product. The difference is larger than price. The unit itself changes.

Parallel’s Search API runs $1–5 per 1,000 calls. Its Task API runs $5 to $2,400 per 1,000 successful runs, with one charge per run rather than per requested output field. Perplexity’s Agent API combines direct model token rates with fees for tools such as web search. Exa Agent offers fixed effort levels from $0.012 to $1 per run, plus a metered mode based on compute and tool use.

The common direction is clear: a query is becoming an input to a larger unit of work. The vendors implement that idea differently, so the rates cannot be compared by lining up three pricing-page numbers.

The measurement set spans two layers

Ranked retrieval provides the shared comparison layer. Agent and task products add state, synthesis, tools, scheduling, and different billing events:

The two layers answer different questions. A retrieval leaderboard cannot choose an agent research system. Parallel’s platform can preserve state and return a research basis. Exa Agent can scale effort and continue a prior run. Perplexity’s Agent API can choose tools and models. AgentCore Web Search stays a retrieval primitive that an agent runtime composes with a model.

The output checks matter as much as the endpoint checks. In one frozen fact pattern, five of 13 Perplexity outputs were incorrect. Parallel’s direct-answer surfaces produced partial and incorrect answers while all ten Task processors completed a separate modes-and-pricing task correctly. These are narrow observations, not accuracy rankings. They prove that HTTP success, tool availability, and answer quality are three separate measurements.

What’s missing

The current evidence does not include a controlled cross-vendor task benchmark. The saved tasks differ in prompts, output schemas, effort settings, and stopping conditions. A fair comparison needs the same research job, the same required evidence, a shared answer rubric, and cost measured through completion. Until then, retrieval and agent-product results remain two evidence lanes.

So what

If you’re choosing a search API for an agent, the vendor comparison you actually need isn’t “who scores highest.” It is three separate questions: which product shape does the vendor sell, what event creates a charge, and what evidence must the system return before the job counts as complete. Get any one wrong and you’re comparing retrieval latency with research quality, or a per-query estimate with a multi-step task bill.

Capital already treats this as infrastructure. Tavily sold for $275M, while Exa, Perplexity, and Parallel reached multi-billion-dollar valuations around the same thesis. The September 11–12 retrieval measurements show who was fastest under one fixed configuration. The product and pricing evidence shows a market defined by more than retrieval.

The open question is whether a shared completed-task benchmark can compare these systems without erasing the product differences that matter most.

Part 1 of the Search for Agents series. Part 2: What actually differs, feature by feature → · Part 3: Search is a pipeline, not a tool → · Part 4: A 200 response does not prove search worked →