TL;DR

• Cheapest per-query at scale: Serper at $0.30-$1.00/1k queries depending on plan, with a 2,500-query free tier per month
• Best value for agent workloads: Brave Search API Data for AI tier - fresh web results plus structured data extraction at $0-$5/mo for low volume, $5/mo for 2k queries/mo on the Pro plan
• Tavily is purpose-built for AI agents with relevance filtering and answer extraction included; $0.008/query on the Research plan
• Exa neural/semantic search starts at $0.001/result but costs compound fast with content retrieval add-ons
• SerpAPI is accurate but expensive at $50/5k queries ($10/1k) - use it for Google result fidelity, not volume
• Bing Web Search API: Microsoft announced end-of-life for March 2025. Do not build new pipelines on it
• Jina Reader and Firecrawl price content extraction separately from search - budget for both if your agent needs to read page content

The Bottom Line

Search is the most underpriced line item in most agent cost models. Teams obsess over LLM token costs and ignore that at 1M agent queries per month, search API bills routinely exceed model bills. The spread across providers is extreme: from $0.30/1k queries at the cheap end to $10/1k at the premium end - a 33x range for roughly equivalent web results.

The complication is that "search" means different things depending on your pipeline. A basic SERP provider returns 10 blue links. An agent-specific provider like Tavily returns ranked, relevance-filtered snippets with an answer synthesis layer. Exa returns semantically matched results from a neural index. Jina Reader and Firecrawl handle the content extraction step - what happens after you get a URL back. These are not directly comparable products, so I've organized pricing around what you're actually building.

For benchmark data on search quality and agent task completion rates, see the AI agent benchmarks leaderboard. For RAG pipeline architecture guidance, see best RAG tools 2026.

Ranked Pricing Table

Sorted by effective cost per 1,000 queries on a mid-volume plan. Where providers charge per request with response count multipliers (e.g., 10 results per query), I've normalized to per-query cost. Prices verified April 19, 2026.

Brave Search pricing is subscription-based, not per-query. Per-1k equivalent shown for the Data for AI $5/mo plan at 2k queries/month = $2.50/1k. Free tier gives 2k/mo at $0.

Cost at Scale

Raw dollar cost at 10k, 100k, and 1M queries per month. Where providers use subscription tiers, I've taken the most cost-effective matching tier.

The Serper vs. Tavily gap is the most striking: at 1M queries, Serper costs roughly $270-$750 vs. Tavily's $8,000. The difference is what you're paying for. Serper gives you raw Google SERP results. Tavily gives you ranked, filtered, agent-ready content. Whether that 10x cost differential is worth it depends on how much post-processing your pipeline does in-house.

Per-Provider Breakdown

Brave Search API

Pricing: Subscription tiers. Free: 2,000 queries/month. Basic: $5/month for 20,000 queries. Starter: $20/month for 100,000 queries. Standard: $40/month for 500,000 queries. Premium: custom pricing.

What you get: Brave's Data for AI plan is specifically built for LLM grounding and RAG. Results include rich structured data - infoboxes, FAQs, news results, location data, and discussions from Reddit/forums. The freshness tier is configurable: you can request results indexed within 24 hours, 7 days, or 30 days. Brave runs its own independent web crawler rather than licensing Google data, which matters for agents that need to avoid a single point of dependency.

Best fit: Agent pipelines needing structured grounding data at predictable monthly costs. The $5/month tier handles 20k queries - more than sufficient for a small production agent. The flat-rate model makes budget forecasting easy.

Gotchas: Coverage is narrower than Google-backed providers. Freshness is configurable but near-real-time, not instant. The free tier's 2k/month cap runs out fast in development if your agent is chatty. Enterprise pricing is contact-sales, so modeling costs above 500k/month requires a call.

Source: Brave Search API

Tavily

Pricing: Free: 1,000 searches/month. Research plan: pay-as-you-go at $0.008/search ($8/1k). Custom enterprise contracts above ~1M/month.

What you get: Tavily is the most agent-native search API in this comparison. It was built from day one for LangChain, LlamaIndex, and similar frameworks - the SDK ships with direct integrations. A single Tavily query returns pre-ranked results with relevance scores, extracted content snippets, and an optional AI-generated answer. That answer synthesis step means you might be able to skip a separate LLM call for simple factual questions. Tavily also includes search depth settings: basic (faster, cheaper) vs. advanced (slower, more sources scraped).

Best fit: Agent loops that query search frequently and need results pre-processed for LLM consumption. If your agent is doing multi-step research, the relevance filtering reduces noise significantly compared to raw SERP results.

Gotchas: $8/1k is expensive if you hit the search step frequently. At 100k searches/month, Tavily costs $800 - compare that to $75 with Serper Pro. If your pipeline can handle raw SERP results with your own post-processing, Tavily's value proposition weakens at scale. The advanced search depth costs more and is not clearly documented in per-request pricing.

Source: Tavily

Exa

Pricing: Free: 1,000 searches/month. Paid plans start at $0.001/result (where each search returns up to 10 results, so $0.01/search at maximum results). Content retrieval (fetching full page text) is an additional $0.001/page. Custom pricing at scale.

What you get: Exa is semantically different from every other provider here. Instead of keyword matching against a traditional web index, Exa uses neural search - meaning you can send natural language queries and get back semantically similar documents rather than keyword-matched pages. This works well for research tasks where you're looking for documents like "academic papers on attention mechanisms" rather than querying with exact keyword phrases. Exa indexes the web but their index is optimized for clean, structured content - academic papers, documentation, news, blogs - rather than the long tail of the web.

Best fit: Research agents, document retrieval pipelines, and RAG systems where semantic similarity matters more than freshness. Not the right tool for "what happened today in AI news."

Gotchas: The per-result pricing looks cheap ($0.001/result) but compounds quickly. A search returning 10 results plus content extraction for each result = $0.01 (search) + $0.01 (10 page fetches) = $0.02 per query. That's $20/1k queries - 20x the headline. Model the full pipeline cost, not just the search query price. Freshness lags behind Google-backed providers; indexed content is typically hours to days old.

Source: Exa

SerpAPI

Pricing: Free: 100 searches/month. Hobby: $50/month for 5,000 searches ($10/1k). Production: $130/month for 15,000 searches ($8.67/1k). Business: $250/month for 30,000 searches ($8.33/1k). Agency: $500/month for 70,000 searches ($7.14/1k).

What you get: SerpAPI scrapes Google, Bing, YouTube, Google Images, Google Shopping, and 25+ other search engines. Results are structured JSON - you get the exact data displayed on the SERP, including ads, organic results, knowledge panels, featured snippets, and rich results. This is the most comprehensive structured SERP data in this comparison. SerpAPI also handles JavaScript rendering for sites that require it.

Best fit: Applications that need accurate Google SERP data for competitive analysis, SEO monitoring, or any use case where result structure matters. Also appropriate when you need access to non-web search engines (YouTube, Shopping, Images) through one API.

Gotchas: At $10/1k on the entry plan, it's the most expensive pure search option. The 100-search free tier is too small for meaningful development testing. Pricing per plan is flat-rate but the jump from Hobby ($50) to Production ($130) doubles spend for a 3x query increase.

Source: SerpAPI

Serper

Pricing: Free: 2,500 searches/month. Developer: $50/year for 50k searches ($1.00/1k). Pro: $50/month for 500k searches ($0.10/1k). Business: $250/month for 3M searches ($0.083/1k).

What you get: Serper returns Google search results via a clean JSON API. Results match what you'd see in Google - organic results, answer boxes, related questions, knowledge graphs. Latency is typically under 100ms. The API is minimal but complete: search by query, location, language, and time range. There is no neural re-ranking, no AI answer synthesis, no content extraction - just Google SERP data at competitive prices.

Best fit: Any pipeline that needs Google results at high volume and can do its own content parsing. The best cost/query ratio of any Google-backed provider. If you're running 100k+ searches per month and your agent processes raw SERP responses, Serper should be your starting point.

Gotchas: No content extraction - you get URLs and snippets, not page content. The 2,500-query free tier is generous for testing but the jump to the annual Developer plan ($50/year, 50k total queries) is limited if you're doing high-volume development. The Pro plan at $0.10/1k is genuinely cheap for production, but only kicks in at $50/month - evaluate whether that minimum makes sense for your volume.

Source: Serper

Perplexity Sonar API

Pricing: sonar (lightweight): $1/1k requests + $1/1M tokens. sonar-pro: $3/1k requests + $2/1M tokens (input) and $3/1M tokens (output). sonar-reasoning: $5/1k requests. sonar-deep-research: $5/1k requests + token costs.

What you get: Perplexity's API is not a raw search API - it's a search-augmented LLM that returns synthesized answers with citations. The model queries the web in real time and returns a structured response with inline source references. This is meaningfully different from SERP APIs: you're getting an LLM-generated answer rather than a result list. The sonar-deep-research model runs multi-step research loops internally before returning a response.

Best fit: Agents that need cited answers to factual questions without building their own search-to-synthesis pipeline. If your agent currently runs: (1) query search API, (2) parse results, (3) send to LLM to synthesize answer - Perplexity's API replaces steps 1-3 in one call.

Gotchas: Per-request cost plus token costs makes this hard to budget predictably. A sonar-pro query returning 2,000 input tokens and 500 output tokens costs $3/1k (request) + $2/1k (input tokens at scale) + $3/1k (output tokens at scale) - at moderate volume, total cost per 1k requests can run $8-15 or more depending on response length. You're also paying for LLM reasoning, not just search. If your agent needs raw search results rather than synthesized answers, this is the wrong abstraction.

Source: Perplexity AI API docs

You.com Search API

Pricing: Starter: $99/month for up to 10,000 queries. Growth: $299/month for up to 50,000 queries. Business: contact sales. Per-query equivalents: $9.90/1k and $5.98/1k respectively.

What you get: You.com offers two API products: Smart (real-time web search with AI summaries and citations) and Research (multi-step research agent that returns comprehensive reports). Smart is the standard grounding tool for agents; Research is positioned for automated report generation. Results include cited URLs, web snippets, and an AI-generated summary. You.com's independent web index supplements Google results on some tiers.

Best fit: Agents that need cited answers with transparent source attribution. The AI summary layer reduces post-processing work, similar to Perplexity. You.com's daily free tier (60 queries/day via sign-up) is useful for prototyping.

Gotchas: Pricing is on the expensive end for what amounts to a search-plus-summarize pipeline. At $99/month for 10k queries, you're paying significantly more than Serper or Brave for high-volume use cases. Enterprise pricing is opaque.

Source: You.com API docs

ValueSerp and ScaleSerp

Pricing (ValueSerp): Free: 100 calls/month. Growth: $19/month for 10k calls ($1.90/1k). Business: $49/month for 50k calls ($0.98/1k). Enterprise: $149/month for 200k calls ($0.75/1k).

Pricing (ScaleSerp): Free: 100 calls/month. Start: $5/month for 2.5k calls ($2.00/1k). Growth: $25/month for 25k calls ($1.00/1k). Business: $75/month for 100k calls ($0.75/1k).

What you get: Both are budget Google SERP providers offering JSON-structured search results at prices competitive with Serper. ValueSerp and ScaleSerp target developers who need Google results without the SerpAPI premium. Feature sets are similar: organic results, answer boxes, related questions, local results, image results. Response format is clean JSON. Both support country and language targeting.

Best fit: High-volume pipelines needing Google SERP data at $1/1k or below. Useful as a fallback or secondary provider alongside Serper.

Gotchas: Less brand recognition and ecosystem integration than Serper or SerpAPI. Reliability and uptime SLAs are not prominently documented. Worth testing against Serper for your specific query mix before committing.

Sources: ValueSerp, ScaleSerp

Linkup

Pricing: Free tier available. Paid plans start at approximately $0.007/search (Research plan equivalent). Pricing is public but changes frequently - check linkup.so for current rates.

What you get: Linkup is an independent search API designed for AI agent grounding. It offers standard and deep search modes. Standard mode returns ranked snippets with source URLs. Deep search performs additional content extraction on top results and returns fuller page content alongside the search results. The deep mode positions it similarly to Tavily.

Best fit: Teams looking for a Tavily alternative with agent-focused features. Worth benchmarking against Tavily on result quality for your specific query domains.

Gotchas: Relatively new entrant with less community usage data than Tavily or SerpAPI. Deep search pricing is not clearly broken out per-mode on the public pricing page. Verify current rates at the source before modeling costs.

Source: Linkup

Bing Web Search API - Deprecation Notice

Microsoft deprecated the Bing Web Search API as part of a broader Azure Cognitive Services restructuring, with end-of-life effective March 2025. If your pipeline currently uses the Bing Search API, migrate now. The API stopped accepting new subscriptions in 2024 and existing contracts were honored through Q1 2025.

Microsoft has not announced a direct replacement under the Azure AI services umbrella. Teams migrating from Bing should evaluate Brave Search API (independent index), Serper (Google-backed), or SerpAPI (multi-engine) based on volume and freshness requirements. The Bing index itself continues to exist and is accessible through the Microsoft Copilot products, but there is no programmatic search API equivalent for developers at this point.

Jina Reader API

Pricing: Free: 1 million tokens/month. Paid: $0.002 per 1,000 tokens. At average page sizes (~5k tokens), that's approximately $0.01 per page fetched ($10/1k pages).

What you get: Jina Reader (r.jina.ai) is not a search API - it's a content extraction service. You send it a URL and it returns clean markdown text suitable for LLM ingestion. It handles JavaScript rendering, PDF extraction, and dynamic content. In an agent pipeline, Jina Reader sits at step 2 after you have URLs from a SERP provider: you extract and clean the content for RAG indexing or direct LLM context.

Best fit: Any agent pipeline that needs to fetch and process web pages. The free million-token tier covers substantial prototyping. In production, Jina Reader pairs well with Serper or Brave for search, with Jina handling the content extraction step.

Gotchas: Rate limits on the free tier are not prominently documented. The per-token pricing model means pages with dense text (research papers, news articles) cost more than lightweight pages. Budget both search queries and content extraction separately when modeling total agent costs.

Source: Jina AI Reader

Firecrawl

Pricing: Free: 500 credits/month (1 credit = 1 page scrape). Hobby: $19/month for 3,000 credits. Standard: $83/month for 100,000 credits. Growth: $333/month for 500,000 credits.

What you get: Firecrawl is a full web scraping and crawling platform with an LLM-friendly output mode. It handles JavaScript rendering, anti-bot detection, and formats output as clean markdown. Unlike Jina Reader's single-URL extraction model, Firecrawl can crawl entire sites, follow links, and return structured data. Firecrawl also has a search endpoint that queries the web and returns scraped full-text content - combining search and extraction in one call.

Best fit: Agents that need to deeply crawl a specific site, extract structured data, or index entire domains for RAG. The combined search-plus-extraction mode eliminates a separate SERP API call, but at significantly higher per-query cost than running Serper + Jina Reader separately.

Gotchas: Credits cover crawling not just searching - a single crawl of a large site can burn significant credits. At $0.083/1k credits on the Growth plan, Firecrawl is more expensive per query than most pure search options. The best case for Firecrawl is deep site indexing, not high-frequency lightweight search.

Source: Firecrawl

Hidden Costs

Content Extraction - The Budget Line Everyone Misses

Most SERP providers return URLs and snippets. Agents often need to read the actual page content - the full article, the product description, the documentation page. That requires a second API call to a content extraction service (Jina Reader, Firecrawl, or your own headless browser fleet).

At typical agent workloads where 20-40% of search results get fully fetched:

At scale, content extraction frequently costs more than search. Budget for both from day one.

Per-Engine vs. Per-Query Pricing

SerpAPI charges per search regardless of which engine you hit - Google, Bing, YouTube each costs one query from your plan allowance. If your agent is querying multiple engines per user request (common in research agents), a single "user search" burns multiple API queries. Model your downstream query fan-out, not just the top-level user request count.

Geographic and Language Targeting Costs

Several providers charge more for searches targeting specific countries or non-English languages. SerpAPI's pricing is uniform across locales. ValueSerp and ScaleSerp apply the same rate regardless of location. Verify whether your target language/country is covered before committing to a provider with potential upcharge.

Volume Cliff Risk

Brave's subscription model creates a cliff at each tier: if your agent runs 2,001 queries in a month on the free tier, you've hit the cap. Query bursts from an agent loop that retries on failure can unexpectedly exhaust monthly quotas. Implement circuit breakers and query counting in your agent pipeline to avoid surprise overages or hard rate limits mid-month.

Free Tier Comparison

Serper and Brave offer the most useful free tiers for real development work. 2,500 queries/month is enough to build and test an agent pipeline. Jina Reader's 1M free tokens is particularly generous - that covers roughly 200 average web pages per day with no billing until you scale.

Which Provider for Which Workload

The table below is an opinionated guide based on what I'd build for each scenario, not just the price sheet.

Price History

• Q1 2025 - Microsoft deprecated Bing Web Search API with end-of-life in March 2025. Pipelines relying on Bing programmatic access were forced to migrate. No direct Microsoft replacement announced.

• 2025 - Serper introduced an annual Developer plan ($50/year for 50k queries), undercutting monthly-billed competitors on a per-query basis for developers with moderate volumes.

• 2024 - Tavily launched its agent-specific API with native LangChain and LlamaIndex integrations, targeting the growing AI agent market directly rather than positioning as a general search tool.

• 2024 - Jina AI launched Reader (r.jina.ai) as a standalone content extraction API, making it easy to integrate URL-to-markdown extraction without running a headless browser.

• 2024 - Brave launched its Data for AI subscription tier specifically for LLM grounding use cases, with structured data outputs and freshness controls.

• 2023-2024 - Exa rebranded from "Metaphor" and expanded from pure academic/blog neural search to broader web coverage, adding more recent content to its index.

The market is bifurcating: commodity SERP providers competing on price (Serper, ValueSerp, ScaleSerp), and agent-native providers competing on result quality and LLM integration (Tavily, Exa, Perplexity). The commodity tier is getting cheaper. The agent-native tier is getting more expensive as providers add more processing layers. Know which tier your pipeline actually needs before defaulting to the agent-native options.

FAQ

Which search API is cheapest for AI agents?

For raw cost, Serper is the cheapest Google-backed option at $0.10/1k queries on the Pro plan. Brave Search API's free tier gives 2,000 queries/month at no cost. If you need agent-specific features like relevance scoring and extracted snippets, Tavily at $8/1k is mid-range. Avoid building cost models on SerpAPI for high-volume agent work - $10/1k adds up fast.

What is the difference between Serper and SerpAPI?

Both return Google SERP results in JSON. SerpAPI supports 25+ search engines (Google, Bing, YouTube, Shopping), handles more edge cases, and has more comprehensive documentation - at $10/1k. Serper focuses on Google and returns clean results at $0.10-$1.00/1k. For most agent use cases that only need Google web results, Serper is the better choice on price.

Should I use Tavily or build my own search pipeline?

If you're already using LangChain or LlamaIndex, Tavily is worth testing - the native integrations reduce integration time significantly. At low volume (under 10k queries/month), the $8/1k cost is manageable. At 100k+ queries/month, the $800+ monthly cost starts to look expensive relative to building Serper + your own post-processing. The break-even point depends on your engineering cost to build and maintain that pipeline.

How do I handle content extraction alongside search?

The cheapest full-pipeline approach is Serper (search) + Jina Reader (content extraction). At 100k queries/month with 30% fetch rate, that's roughly $27 + $300 = $327. Compare that to Firecrawl's combined search-and-crawl at $1,500+ for equivalent volume. For high-volume pure search-and-read pipelines, the two-provider approach wins on cost.

Is Bing Web Search API still available?

No. Microsoft ended the Bing Web Search API in March 2025. Do not start new projects on it. Existing integrations should migrate to Brave Search, Serper, or SerpAPI.

What's the best free search API for testing?

Serper's 2,500/month free tier and Brave's 2,000/month free tier are the most practical for real testing. Jina Reader's 1M tokens/month is the most generous for content extraction testing. Tavily and Exa each give 1,000 free queries/month - enough to prototype but not to load-test.

Does Perplexity Sonar replace a separate search API?

For simple factual Q&A agents, yes - Perplexity Sonar returns a synthesized answer with citations and you skip the separate search-to-LLM step. For agents that need to process raw search results (filter by date, apply custom ranking, extract structured data), a raw SERP API plus your own LLM gives more control. Perplexity's per-request plus token pricing also makes cost modeling harder than a flat per-query rate.

Sources:

• Brave Search API
• Tavily
• Exa Pricing
• SerpAPI Pricing
• Serper
• Perplexity AI API Documentation
• You.com API Documentation
• ValueSerp
• ScaleSerp
• Linkup
• Jina AI Reader
• Firecrawl Pricing

Also see: Embedding Models Pricing 2026 and LLM API Pricing Comparison.