Web Search APIs for AI Agents Compared: Six Options Ranked
No single web search API for AI agents wins on every axis. On our weighted scorecard, Brave Search API ranks first at 74.0 out of 100, narrowly ahead of Exa at 73.5 and Parallel at 71.0. Parallel and Perplexity list the lowest per-search price, $1 per 1,000 fast requests, as of 18 November 2026.
A web search API for AI agents is a service that takes a text query from a model and returns ranked web results, usually with snippets or page text, in a format an LLM can read. Choosing one comes down to price per search, how much of the page you get back, who owns the index and what the vendor does with your queries. This guide compares six of them on those points.
On our weighted scorecard, Brave Search API ranks first at 74.0 out of 100, with Exa at 73.5 and Parallel at 71.0, so the top three are close. Every figure below was read from the vendor's own page on 18 November 2026.
How We Ranked These Web Search APIs
This is an editorial synthesis. The ratings are based on documented features and published pricing, checked on 18 November 2026. We did not run benchmarks on any of these products, and we make no claim about result quality or answer accuracy beyond what vendors document. Where a vendor publishes no figure for a criterion, that entry takes the lowest documented rating of 4 on that criterion, never a guess.
Each entry is rated from 0 to 10 on seven criteria. The weighted score is the sum of weight times rating divided by 10, so the maximum is 100. The weights are the same for every entry.
| Criterion | Weight | What earns a high rating |
|---|---|---|
| --- | --- | --- |
| Cost per 1,000 typical agent searches | 25 | Low list price for a 10-result search |
| Pricing transparency | 15 | One clear price list, few contradictions |
| Result freshness and index independence | 15 | Documented own index, no reseller fallback |
| Content extraction for LLMs | 15 | Page text, extraction and crawl endpoints, citations |
| Free tier and ease of start | 10 | A recurring free allowance, no sales call |
| Rate limits and scale ceiling | 10 | Published numeric limits that scale up |
| Data retention and privacy options | 10 | Documented zero data retention or regional options |
No vendor paid for inclusion, position or a better rating, and none has reviewed this page. Rank is determined only by the scores below.
Scorecard: Which Web Search API Ranks Highest?
The table shows each rating (0 to 10) in criterion order: cost, transparency, index, extraction, free tier, rate limits, privacy.
| Rank | API | Cost | Transp. | Index | Extract | Free | Limits | Privacy | Score |
|---|---|---|---|---|---|---|---|---|---|
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Brave Search API | 7 | 7 | 10 | 6 | 6 | 8 | 8 | 74.0 |
| 2 | Exa | 7 | 8 | 8 | 8 | 8 | 5 | 7 | 73.5 |
| 3 | Parallel | 9 | 5 | 4 | 8 | 8 | 8 | 7 | 71.0 |
| 4 | Firecrawl | 6 | 7 | 4 | 10 | 6 | 9 | 4 | 65.5 |
| 5 | Tavily | 6 | 6 | 4 | 8 | 8 | 4 | 4 | 58.0 |
| 6 | Perplexity Search API | 8 | 7 | 4 | 4 | 3 | 4 | 4 | 53.5 |
Arithmetic for the top entry: Brave is 25 x 0.7 = 17.5, plus 15 x 0.7 = 10.5, plus 15 x 1.0 = 15, plus 15 x 0.6 = 9, plus 10 x 0.6 = 6, plus 10 x 0.8 = 8, plus 10 x 0.8 = 8, which totals 74.0. A rating of 4 on index independence means the pages checked did not document where results come from, not that the vendor lacks an index.
1. Brave Search API — Best for an independent index
Best for: teams that want results from a vendor-owned index and documented rate limits.
Brave states that every result comes from its own index, with no fallback to Google or Bing. Its crawler is supplemented by opt-in signals from Web Discovery Project users. That is the clearest index-ownership statement on any page in this comparison, and it earns the top index rating.
- Newest version: Brave publishes no dated changelog that we could find on the pages checked on 18 November 2026. The plan page lists an LLM Context endpoint for agentic search, an Answers endpoint, and web, image, video, news and places search.
- Pricing: Search is $5 per 1,000 requests. Answers is $4 per 1,000 requests plus $5 per million input and output tokens, as of 18 November 2026.
- Limits: Search is rated at 50 queries per second and Answers at 2 queries per second. Enterprise is custom.
- Standout: Brave describes its zero data retention as architectural, with queries staying on infrastructure it runs end to end. It lists full-funnel zero data retention as an Enterprise feature.
Limitations: the page contradicts itself on free credit. The hero text says "Up to $20 in free calls every month", while both plan cards list $5, so only $5 is safe to budget. Extraction is thinner than at the crawl-focused vendors, and no dated release notes means we cannot name a newest version.
2. Exa — Best for tiered search depth
Best for: agents that need to trade latency and cost against depth, search by search.
Exa sells search by type, from a cheap lite type to deep search, and publishes one table for all of them. Its changelog is dated to the day, which makes the freshness claim easy to check.
- Newest version: "Exa Search in the EU", 13 November 2026. Before it came Monitors v2 and Exa Podcasts on 29 October 2026 and the
litesearch type on 19 October 2026. - Pricing:
liteis $4 per 1,000 requests,auto$7,deep$12 anddeep-lite$12. Results beyond 10 cost $1 per 1,000 results and AI page summaries $1 per 1,000 pages. Contents is $1 per 1,000 pages per content type. - Free tier: $10 in credits on signup, resetting to $10 on the first of each month, which Exa says covers up to 2,500 Lite searches.
- Standout: zero data retention is included with Enterprise and can be added by team owners from the dashboard. The annual self-serve Enterprise plan is $12,000 upfront and lists 25 queries per second.
Limitations: the page says "no minimum spend" while also selling a $12,000 annual plan. No queries-per-second limit is stated for Free or pay as you go.
3. Parallel — Best for lowest list price
Best for: cost-sensitive agents that run many searches and can use a fast mode.
Parallel lists the lowest per-search price in this guide. Its Search API costs $1 per 1,000 requests in turbo and fast modes and $5 per 1,000 in basic and advanced modes, per its pricing documentation. Each request includes 10 results.
- Newest version: Images API, 16 November 2026. Parallel MCP 2.0 followed on 5 November, then a Usage Data API on 4 November.
- Pricing: Extract is $1 per 1,000 URLs, Decisions $1 per 1,000 requests and Task from $5 per 1,000 successful runs, as of 18 November 2026. Scale is $1,000 a month for $1,000 in credits.
- Free tier: Build lists $5 free monthly with a card on file, plus up to $80 at signup and up to 5,000 free requests a month.
- Limits: Build is 600 search requests per minute and Scale 1,500, excluding Turbo.
Limitations: the pricing page and its FAQ conflict on six points. Extract is priced per 1,000 URLs in the table and per 1,000 results in the FAQ, the Build free credit requires a card on the plan table but not in the FAQ, and the FAQ says "Launch isn't an account plan" while the page shows it as a plan column. Zero data retention is hybrid and needs Scale plus a signed agreement.
4. Firecrawl — Best for search plus scraping
Best for: pipelines that search, then scrape or crawl the pages found.
Firecrawl sells one credit pool across search, scrape, map, crawl and agent endpoints, so the same plan covers finding pages and turning them into markdown. Search costs 2 credits per 10 results, per its billing documentation.
- Newest version: UltraSearch, 17 November 2026, adding data on 250 million-plus people, 75 million-plus companies and 150,000-plus stocks. Before that, MCP v3 on 12 November and Alexandria company coverage on 13 November.
- Pricing: Free is $0 for 1,000 credits, Hobby $19 a month for 5,000, Standard $99 for 100,000, Growth $399 for 500,000 and Scale $749 for 1,000,000 credits, as of 18 November 2026.
- Limits: the most detailed table here. Search is rated 10 per minute on Free, 100 on Hobby, 500 on Standard and 10,000 on Scale.
- Standout: extra credits are sold in $5 increments, at $0.005 per credit on Hobby falling to $0.001 on Scale.
Limitations: the page calls the concurrency limit "concurrent requests" in one place and "concurrent browsers" in another. Pay as you go credits cost more than plan credits, about $0.0025 against $0.00099 on Standard. The page does not document zero data retention or where search results originate.
5. Tavily — Best for a simple credit model
Best for: developers who want search, extract, map and crawl behind one small API and a free start.
Tavily prices in credits. A basic search costs 1 credit and an advanced search 2, and every 5 successful URL extractions cost 1 credit on the basic setting, per its credits documentation.
- Newest version: Tavily CLI 1.0 on npm, November 2026. The changelog gives only the month. It also lists per-organization rate limits in November 2026 and parallel CLI searches in October 2026.
- Pricing: Researcher is free with 1,000 credits a month. Pay as you go is $0.008 per credit. Project is $30 a month for 4,000 credits, about $0.0075 per credit, as of 18 November 2026.
- Limits: the pricing page lists "higher rate limits" for Project and "custom" for Enterprise, with no numbers. The changelog says limits apply per organization and are shared across API keys.
- Standout: a startup program offers 100,000 free credits.
Limitations: the FAQ says requests stop when credits run out, while pay as you go is described as paying only for use. The pages checked do not document zero data retention or index ownership.
6. Perplexity Search API — Best for a cheap fast tier
Best for: teams already on Perplexity that want a low-cost raw search endpoint.
Perplexity's Search API charges per successful request, with up to five queries in one request counting as one unit. It adds no token charges for search, and failed or rate-limited requests are not billed.
- Newest version: Decisions API models pplx-decider-v1.3 and pplx-decider-flash-v1.3, November 2026. The changelog gives only the month.
- Pricing: standard Search is $5.00 per 1,000 requests and fast search ($1.00 per 1,000) is selected with
search_type: "fast", as of 18 November 2026. - Limits: the pricing page gives none and links to a separate rate-limit page.
- Standout: multi-query requests billed once make batching cheap.
Limitations: the page conflicts with itself on standard search. The Agent API web_search tool is $0.0025 per call, or $2.50 per 1,000, while the Search API standard price is $5.00 per 1,000. The pages checked list no free tier, no extraction endpoint and no data retention terms.
Summary Comparison Table
| API | Cheapest listed search | Free allowance | Newest dated release | Zero data retention documented |
|---|---|---|---|---|
| --- | --- | --- | --- | --- |
| Brave | $5 per 1,000 | $5 a month on plan cards | No dated changelog found | Architectural, full-funnel on Enterprise |
| Exa | $4 per 1,000 (Lite) | $10 a month | 13 November 2026 | Enterprise |
| Parallel | $1 per 1,000 (fast) | $5 a month, up to 5,000 requests | 16 November 2026 | Scale, hybrid |
| Firecrawl | 2 credits per search | 1,000 credits on Free | 17 November 2026 | Not documented |
| Tavily | 1 credit per search | 1,000 credits a month | November 2026 (no day) | Not documented |
| Perplexity | $1 per 1,000 (fast) | None listed | November 2026 (no day) | Not documented |
Which Web Search API Is Cheapest for AI Agents?
By list price, Parallel's fast mode and Perplexity's fast search are cheapest at $1 per 1,000 requests, as of 18 November 2026. Exa Lite follows at $4, Brave and the standard Parallel and Perplexity modes are $5, and Tavily is about $7.50 to $8. The worked example below shows the arithmetic.
Assumptions: 1,000 searches, 10 results each, no free credits applied, list prices from each vendor's page, search only.
| API and mode | Arithmetic | Cost per 1,000 searches |
|---|---|---|
| --- | --- | --- |
Parallel fast | 1,000 x $0.001 | $1.00 |
| Perplexity fast search | 1,000 x $0.001 | $1.00 |
| Firecrawl Standard | 2,000 credits x ($99 / 100,000) | $1.98 |
Exa lite | 1,000 x $0.004 | $4.00 |
| Brave Search | 1,000 x $0.005 | $5.00 |
Parallel basic | 1,000 x $0.005 | $5.00 |
| Perplexity standard | 1,000 x $0.005 | $5.00 |
Exa auto | 1,000 x $0.007 | $7.00 |
| Tavily Project | 1,000 credits x ($30 / 4,000) | $7.50 |
| Firecrawl Hobby | 2,000 credits x ($19 / 5,000) | $7.60 |
| Tavily pay as you go | 1,000 credits x $0.008 | $8.00 |
Firecrawl's figure depends on the plan you buy, since Standard needs $99 a month, and its pay as you go credits at $0.005 on Hobby would make the same 1,000 searches $10.00. Plan credits you do not use are a real cost, so compare monthly volume as well.
Now add page text. Assume 1,000 searches plus extraction of 5 pages per search, or 5,000 pages, where each vendor lists a price:
- Exa:
litesearch $4.00 plus contents 5,000 x $0.001 = $5.00, so $9.00. - Parallel:
fastsearch $1.00 plus Extract 5,000 x $0.001 = $5.00, so $6.00. - Tavily: basic search $8.00 plus 5,000 / 5 = 1,000 credits x $0.008 = $8.00, so $16.00.
We leave Brave, Perplexity and Firecrawl out of this second example. The Brave and Perplexity pages checked list no page extraction price, and Firecrawl's scrape credit rate was not read, so we do not estimate it.
How Do Web Search APIs Price Their Plans?
Web search APIs price in three ways: per request (Exa, Brave, Parallel, Perplexity), in credits that vary by endpoint (Tavily, Firecrawl), or in monthly tiers with included credits (Firecrawl, Tavily Project, Parallel Scale). Extras such as extra results, page contents and summaries are usually billed separately, so a "per search" price is rarely the full bill.
Three details change real cost. First, extra results: Exa and Parallel both charge $1 per 1,000 results beyond 10. Second, depth modes: Exa moves from $4 to $12 per 1,000 as you move from Lite to deep, and Parallel from $1 to $5 between its fast and basic modes. Third, expiry: Parallel Scale monthly credits expire at the end of each paid month, and Firecrawl publishes its rollover rules in its billing documentation.
A rule of thumb for budgeting is to take your expected monthly searches, multiply by the per-1,000 price above, add extraction, then compare with the plan price. A flat plan wins only if you use most of the credits.
Brave vs Exa: Which Should You Choose?
Choose Brave if you want an independent index, documented rate limits and a plain $5 per 1,000 price. Choose Exa if you want several search depths from $4 to $12 per 1,000, a contents endpoint and a day-dated changelog. They score 74.0 and 73.5, so the gap is smaller than any rating's uncertainty.
Brave's edge on this scorecard is index ownership (10 against 8) and published rate limits (8 against 5). Exa's edge is extraction (8 against 6), free tier (8 against 6) and pricing transparency (8 against 7). If your agent needs only snippets, Brave's flat price is simpler. If it needs page text and recent content, Exa's Contents price of $1 per 1,000 pages is easy to model.
Which Web Search API Should You Choose?
For an agent that needs an independent index and predictable limits: Brave Search API. It states its own index, publishes 50 queries per second for Search and prices flatly at $5 per 1,000.
For flexible depth and day-dated releases: Exa. Pick lite for cheap lookups and deep for harder ones, and check its 13 November 2026 EU availability if data location matters.
For the lowest list cost: Parallel fast at $1 per 1,000, with its page contradictions read first. Perplexity's fast tier matches the price but adds no extraction or free tier on the pages checked.
For search followed by scraping: Firecrawl, where one credit pool covers search, scrape, map and crawl. Budget the plan, not the credit rate.
For a quick free start: Tavily's 1,000 free credits a month, or Exa's $10 monthly credit, need no sales call.
For retrieval pipelines: pair any of these with the steps in our RAG pipeline tutorial, and if answers go wrong, use our RAG debugging guide.
Which Web Search API Works Best With MCP and Agent Frameworks?
Several vendors now ship MCP servers, which let an agent call search as a tool. Parallel released MCP 2.0 on 5 November 2026, replacing separate Search and Task servers with one server offering web search, fetch, decisions and research tools. Firecrawl released MCP v3 on 12 November 2026.
If you are new to the protocol, our MCP guide explains how servers and clients connect. For wider context on which agents use web tools, see our autonomous AI agents guide. If you route calls through a proxy layer, our LLM gateways comparison covers that side.
Also Considered, Not Scored
- Serper: its
serper.dev/pricingpage returned HTTP 404 when fetched on 18 November 2026, so we could not read a price and did not score it. - Jina Reader: its page shows Free, Paid and Premium tiers with rate limits such as 100 requests per minute on Free, but no dollar amounts, so we could not compute cost per 1,000 searches.
We also do not rank consumer AI search products here. Those are covered in our AI search engines comparison, which has a different purpose from a developer API.
Contradictions Found on Vendor Pages
- Brave: "Up to $20 in free calls every month" in the hero text, against $5 on both plan cards, at the vendor's pricing page.
- Exa: "no minimum spend" against a $12,000 upfront annual plan, at the vendor's pricing page.
- Parallel: six conflicts between table and FAQ, including Extract per URL versus per result and Launch as a plan versus "isn't an account plan", at the vendor's pricing page.
- Firecrawl: concurrency described as "requests" and as "browsers", at the vendor's pricing page.
- Tavily: requests stop at zero credits versus pay as you go, on the vendor's pricing page.
- Perplexity: Agent API
web_searchat $2.50 per 1,000 against Search API standard at $5.00 per 1,000, at the vendor's pricing page.
Where figures conflict we used the more conservative number in scoring and name both here.
Conclusion
The top three here are within 3 points, which is a reason to test the leader against your own queries before committing. Brave leads on index ownership and limits, Exa on tiered depth and clean pricing, and Parallel on cost. The cost ranking changes sharply once you add page extraction, so model that, not just the search price.
Prices and changelogs move weekly in this category. We checked each page on 18 November 2026, and you should confirm current figures with each vendor before you commit.
This is an editorial synthesis of documentation, pricing pages and changelogs, not hands-on benchmarking; see our [methodology](/methodology). Verify current details with each vendor.
Sources
Primary sources checked on 18 November 2026:
- Brave Search API plans (brave.com)
- Exa pricing (exa.ai)
- Exa changelog (docs.exa.ai)
- Parallel pricing (parallel.ai)
- Parallel Search API pricing documentation (docs.parallel.ai)
- Parallel changelog (docs.parallel.ai)
- Firecrawl pricing (firecrawl.dev)
- Firecrawl changelog (firecrawl.dev)
- Tavily pricing (tavily.com)
- Tavily API credits (docs.tavily.com)
- Tavily changelog (docs.tavily.com)
- Perplexity API pricing (docs.perplexity.ai)
- Perplexity API changelog (docs.perplexity.ai)
- Jina Reader (jina.ai)
Key Takeaways
- Brave Search API scores 74.0 out of 100, Exa 73.5 and Parallel 71.0, so the top three sit within 3 points on this scorecard.
- Parallel and Perplexity both list $1 per 1,000 requests for their fast search modes, against $5 per 1,000 for their standard modes, as of 18 November 2026.
- On a 1,000-search workload with 10 results each, list cost ranges from $1.00 (Parallel fast) to $8.00 (Tavily pay as you go).
- Firecrawl's Standard plan works out to about $1.98 per 1,000 searches, but only once you commit to $99 a month for 100,000 credits.
- Brave's own page promises up to $20 in free monthly calls in its hero text while its plan cards list $5, so the guaranteed amount is $5.
- Parallel's pricing page and FAQ conflict on six points, including whether Launch is an account plan, so read its FAQ before budgeting.
- Zero data retention is documented on Exa Enterprise, Parallel Scale and Brave Enterprise, and not on the Tavily, Firecrawl or Perplexity pages checked.
Frequently Asked Questions
What is the best web search API for AI agents?
On our weighted scorecard, Brave Search API ranks first at 74.0 out of 100, followed closely by Exa at 73.5 and Parallel at 71.0. Brave leads on index independence and documented rate limits. The best choice still depends on whether you need raw results, full-page extraction or the lowest cost per search.
What is the cheapest web search API for LLMs?
By list price, Parallel's fast mode and Perplexity's fast search are cheapest at $1 per 1,000 requests, as of 18 November 2026. Exa Lite is $4, Brave is $5 and Tavily is $8 on pay as you go. Firecrawl can reach about $1.98 per 1,000 searches, but only on a $99 monthly plan.
Is there a free web search API for AI agents?
Several have free allowances, not unlimited use. Tavily gives 1,000 credits a month, Exa gives $10 of credit each month, Firecrawl's Free plan includes 1,000 credits, Brave lists $5 a month, and Parallel's Build plan lists up to 5,000 free requests a month with a card on file. Perplexity's page lists none.
Do I need a search API if my LLM already has built-in web search?
Often yes. A standalone API lets you choose the provider, control the number of results, cap spend and keep search logic portable across models. Built-in search is simpler for prototypes. If you route several models, compare it with the gateway options in our LLM gateways guide before deciding.
Which search APIs offer zero data retention?
Among the pages checked on 18 November 2026, Exa includes zero data retention on Enterprise, Parallel includes a hybrid version on Scale with a signed agreement, and Brave describes its approach as architectural and lists full-funnel zero data retention as an Enterprise feature. Tavily, Firecrawl and Perplexity pricing pages did not document it.
Why do search API prices look so different?
The billing units differ. Tavily and Firecrawl charge in credits, Exa, Brave, Parallel and Perplexity charge per request, and some add per-page or per-token fees. Comparing them needs one workload and shown arithmetic, which is why this guide converts every vendor to cost per 1,000 searches.
Reviewed by
Carlos Montoya
Growth CTO, Open English
Reviewed for technical accuracy on November 18, 2026
Carlos Montoya is a technologist with two decades of experience building and leading software products, including SaaS, AI-driven applications and blockchain products, and has served as CTO several times. He is Growth CTO at Open English, co-founder and CTO of TeamClass, and a resident technologist at Entrepreneurs Roundtable Accelerator. He was previously co-founder and CTO of Metatate, a data-governance layer exposed through the Model Context Protocol, and VP of Engineering at accel-EQ.