Tavily
Tavily is a web access API for AI agents with a genuinely usable free tier (1,000 credits/month, no card), bundled full-page text in search results, and an official MCP server. But its Research endpoint is dynamically priced at 4–250 credits per request, the headline 1,000 RPM applies only to paid Production keys, entity and academic retrieval lag specialized indexes by large margins, and query data may be used for training unless Zero Data Retention or the data-use setting is changed.
What is Tavily?
Tavily is a web access API for AI agents, built around a single credit-based unit that covers search, extraction, mapping, crawling, and multi-step research. Its official positioning is that it provides the web layer for agents—the retrieval and content-extraction functions that an LLM needs before it can reason or act. The product surface includes Search (basic and advanced), Extract (basic and advanced), Map, Crawl, and Research, plus an official hosted MCP server that exposes the same tools to MCP-compatible clients such as Claude, ChatGPT, and coding agents.
The company was acquired by Nebius, an Amsterdam-based AI cloud company, in February 2026, and Tavily is being integrated into Nebius's cloud platform. That matters for buyers evaluating vendor stability and ecosystem fit: Tavily remains an independent product, but its roadmap and infrastructure are now aligned with Nebius's cloud strategy, which is a different ownership posture from independent search API laboratories such as Exa.
Tavily's pricing model is simpler than per-endpoint pricing in one respect and more complex in another. The simple part is the credit: a basic search is 1 credit, advanced is 2, basic Extract is 1 credit per 5 successful URLs, and Map is 1 credit per 10 pages. A team can model search and extraction costs from these constants. The complex part is Research, which is dynamically priced between 4 and 250 credits per request depending on the model (mini or pro) and the task. A single pro Research request at the maximum costs a quarter of the free monthly allowance. This is not hidden, but it is a range rather than a fixed price, and it makes Research the one Tavily endpoint whose cost cannot be computed from the pricing page alone.
The free tier is one of Tavily's strongest features and one of its most misunderstood. It provides 1,000 credits per month with no credit card, resets monthly, and is enough for roughly 1,000 basic searches or 500 advanced searches. Keyless Search and Extract are available without an account, which lowers the barrier to a first test to essentially zero. But the free tier is not a production tier: Production API keys require an active paid plan or Pay as You Go enabled, and the published 1,000 RPM rate limit applies to Production keys, not to the free Researcher plan. The free tier's practical throughput is lower and less consistently documented, and flex-tier requests that require live content fetching can be refused under load with HTTP 434.
How does Tavily work?
A developer sends a request to one of the endpoints. Search takes a query and returns ranked results with content; basic search returns fewer results and a shorter content payload, advanced search returns more results and more content at 2 credits. Extract takes a list of URLs and returns page content at 1 credit per 5 successful extractions (basic) or 2 credits per 5 (advanced). Map enumerates a site's pages at 1 credit per 10 pages returned. Crawl combines mapping and extraction, so its cost is the sum of both operations. Research runs a multi-step investigation and returns a cited report, with the cost bounded between 4 and 250 credits.
The official hosted MCP server exposes these endpoints to MCP clients. A user connects the client to Tavily's MCP endpoint, supplies an API key, and the client can then call search, extract, map, crawl, and research tools during a conversation or agent workflow. The MCP server consumes credits from the same account balance as direct API calls, so MCP usage is metered identically. Third-party MCP wrappers also exist—for example, an Apify-hosted Tavily Search MCP server priced from $11.99 per 1,000 searches—but those are separate products with their own pricing and should not be confused with Tavily's official MCP server.
The output is model-ready: search results include titles, URLs, content snippets or full text depending on the search depth, and Tavily's documentation emphasizes that the API is designed to return content an LLM can consume without a separate scraping step. Extract returns raw page content, and the caller is responsible for filtering it to the relevant passages before sending it to the model. Research returns a synthesized, cited response, which is the closest Tavily comes to an "answer endpoint" rather than a retrieval endpoint.
Use Cases
- General web search for AI agents and RAG pipelines where the query is a topic or question rather than a specific person, company, or paper
- Retrieval-augmented generation where bundled full-page text simplifies the pipeline, accepting the downstream token cost of sending more content to the model
- Site mapping and crawling for content ingestion, with predictable per-page credit costs
- Multi-step research with cited reports, where the 4–250 credit range is acceptable because the alternative is multiple search and extraction calls
- Prototyping and early-stage agent development where the 1,000-credit free month and keyless Search/Extract remove friction
- Teams already using Nebius infrastructure that want vendor consolidation across AI cloud and web retrieval
- MCP-based agent workflows that need a hosted search and extraction server with a single API key
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| Researcher (Free) | $0 | per month | 1,000 API credits per month. No credit card required. Basic search costs 1 credit; advanced search costs 2 credits. Basic Extract costs 1 credit per 5 successful URL extractions. Map costs 1 credit per 10 pages. Research requests consume 4–250 credits each depending on model and task. The plan resets on the 1st of each month; unused credits do not roll over. Keyless Search and Extract are available without an account, subject to separate per-IP limits. | — |
| Project | $30 | per month | 4,000 credits per month ($0.0075 per credit). The lowest paid tier. Removes the free-tier rate ceiling and enables Production API keys. Research credit consumption remains 4–250 credits per request. | — |
| Bootstrap | $100 | per month | 15,000 credits per month ($0.0067 per credit). | — |
| Startup | $220 | per month | 38,000 credits per month ($0.0058 per credit). | — |
| Growth | $500 | per month | 100,000 credits per month ($0.005 per credit). The highest self-serve plan. | — |
| Pay as you go | $0.008 | per credit | Available once a plan's credit limit is reached. Charges $0.008 per credit for overage. No subscription required on this path alone, but Production API keys require either an active paid plan or PAYGO enabled. | — |
| Enterprise | Custom | annual | Custom volume pricing, custom QPS limits, Zero Data Retention, SLA-backed support, and pricing above 1M calls per month. Arranged through sales. | — |
Common Limitations
- Research pricing is dynamic and can consume an entire free month in one request · Tavily's Search and Extract pricing is predictable: 1 credit for basic search, 2 for advanced, 1 credit per 5 URL extractions. Research is not. The documentation states that Research requests cost between 4 and 250 credits each, with the mini model capped at 110 and the pro model at 250. A single pro Research request at the maximum consumes a quarter of a 1,000-credit Free month, or 6.25% of a $500 Growth plan's 100,000-credit allowance. Because the price is a range rather than a fixed number, a developer cannot know the cost of a Research request in advance from the documentation alone; the actual charge depends on the model and the task's internal complexity. This is a material difference from the rest of Tavily's pricing, where the credit cost is a published constant. Teams using Research at volume should instrument per-request credit consumption rather than budgeting from the headline rate.
- The published 1,000 RPM rate limit is a paid-tier figure, not a free-tier guarantee · Tavily's rate-limit documentation lists Development keys at 100 requests per minute and Production keys at 1,000 RPM. That table is accurate, but it omits a crucial condition: Production keys require an active paid plan or Pay as You Go enabled. The Free Researcher plan does not qualify for a Production key. Third-party sources describe the free tier's practical limit as approximately 1 request per second, or about 5 requests per minute, with bursts above the limit returning HTTP 429. There is also a 'flex tier' behavior on non-paying plans: search requests that set max_age_hours are served under a flex tier and can be refused with HTTP 434 when Tavily is under heavy load. The net effect is that the 1,000 RPM figure that appears in comparisons is not available to free-tier users, and the free tier's actual throughput is materially lower and less consistently documented. Teams planning concurrent agent workloads should not assume 1,000 RPM applies before upgrading.
- No dedicated entity index means people, company, and academic retrieval lag specialized competitors · Tavily relies on its general web index for all query types. It does not maintain separate indexes for people, companies, or academic publications. On Exa's published benchmarks—which Exa runs, so they should be read with that caveat—Tavily's top-result recall is 40.5% for people profiles versus Exa's 75.5%, 61.3% for company search versus 81.5%, and 31.8% for academic publication retrieval versus 63.3%. Exa's comparison page attributes this gap to a structural difference (specialized indexes versus general web retrieval), not a quality difference, and that attribution is plausible: general web retrieval for a named person often surfaces job boards, conference pages, or secondary summaries rather than the profile itself. The practical consequence is that Tavily is a reasonable default for general web search and page content, but a weaker choice when the primary task is finding a specific person, company, or paper. Teams with entity-heavy workloads should evaluate both APIs on their own queries rather than assuming the gap does or does not apply to their use case.
- Independent benchmark results are mixed, and vendor claims should be read with the test set in mind · Tavily's own evaluation repository reports 93.3% on SimpleQA, ahead of Perplexity (85.9%), Google via Serper (82.2%), Brave (76.1%), and Exa (71.2%). That result is real, but it is vendor-run and SimpleQA is a specific factual-QA dataset. Independent evaluations point in different directions: the you.com evaluation framework found Exa at 95.40% and Tavily at 88.53% on SimpleQA, and another Tavily configuration scored 66.14%. The AIMultiple agentic search benchmark of 100 AI/LLM queries scored Tavily at 13.67, behind Brave (14.89), Firecrawl (14.58), and Exa (14.39), with Brave being the only competitor to consistently outperform Tavily by a statistically significant margin. The honest synthesis is that Tavily is competitive on general web search and factual QA, but it is not the benchmark leader across every query type, and its strongest published numbers come from its own test configurations. Buyers should run a representative query set rather than selecting on a single benchmark score.
- Privacy defaults permit training on research inputs unless Zero Data Retention is enabled · Tavily's platform terms permit Tavily and its AI providers to train models on input and output from its AI research features. The 'Allow Use of Query Data' setting in account settings controls whether query data is used to improve services; when ON, query data may be used; when OFF, it is not stored or used for improvements. Zero Data Retention is a separate capability that must be enabled per account, and custom QPS, SLA-backed support, and pricing above 1M calls are arranged through sales. The default posture is therefore not zero-retention: a free or self-serve user who does not change the setting is on a path where query data may be used for service improvement and research features may be used for model training. For teams processing sensitive queries, the configuration change is mandatory, and the Enterprise route is the documented path for contractual ZDR guarantees.
- Bundled full-page text is a token-cost trade-off, not an unmixed advantage · Tavily's search response includes raw page content for results, which Exa's comparison page acknowledges as a genuine Tavily advantage for teams that want full text without a separate extraction call. But raw page text is not the same as query-relevant text. Exa's critique—that its own highlights are selected against the query, returning 500 characters that match the question versus 8,000 characters of static page content—points to a real cost difference at the model layer: the caller pays for every token sent to the LLM, and static full-page text includes navigation, boilerplate, and passages irrelevant to the query. Tavily's Extract endpoint returns static page content by design, leaving relevance filtering to the caller. For a RAG pipeline at scale, the token bill for processing bundled raw text can exceed the search API bill. Bundled text simplifies integration; it does not eliminate the need to budget for downstream token consumption.
Key Features
- Search endpoint with basic (1 credit) and advanced (2 credits) depth
- Extract endpoint for page content at 1–2 credits per 5 successful URLs
- Map endpoint for site enumeration at 1–2 credits per 10 pages
- Crawl endpoint combining mapping and extraction
- Research endpoint with dynamically priced multi-step investigations (4–250 credits)
Tavily vs. Exa and Brave Search API
This comparison uses official pricing pages, documentation, and independently published benchmark data available as of November 2026. It focuses on the structural differences that determine retrieval accuracy by task type, throughput ceiling, content bundling, and privacy defaults—not on declaring a universal winner. Tavily is a general web search API with bundled full-page text and a unified credit unit; Exa is a semantic search API with specialized entity and academic indexes; Brave Search API is an independent index built for high-throughput keyword retrieval. The three overlap at the search endpoint but diverge on what they index, what they bundle into the search price, and how much throughput the entry tier actually provides.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| Tavily Official credits and pricing Official rate limits Query data setting (official help) | A web access API for AI agents with Search, Extract, Map, Crawl, and Research endpoints, a unified credit model, bundled full-page text, and an official hosted MCP server. | The free tier is genuinely usable for prototyping (1,000 credits/month, no card) and search results include raw page text without a separate extraction call. But Research is dynamically priced at 4–250 credits per request, the 1,000 RPM limit requires a paid Production key, and entity/academic retrieval lags Exa's specialized indexes by large margins (40.5% vs. 75.5% on people R@1, 31.8% vs. 63.3% on academic R@1, per Exa's benchmarks). Query data may be used for training unless ZDR or the data-use setting is changed. Tavily was acquired by Nebius in February 2026. | General web search and RAG pipelines where bundled full-page text simplifies the workflow, and teams that value a unified credit unit across search, extraction, mapping, and crawling—especially those already in the Nebius ecosystem or needing a zero-friction free tier for prototyping. | Free Researcher: 1,000 credits/month, no card, monthly reset. Project $30/mo: 4,000 credits ($0.0075/credit). Bootstrap $100/mo: 15,000 credits. Startup $220/mo: 38,000 credits. Growth $500/mo: 100,000 credits. Pay as you go: $0.008/credit. Research: 4–250 credits per request. Enterprise: custom, ZDR available. |
| Exa Exa vs Tavily benchmarks (Exa-run; read with vendor bias) Independent agentic benchmark (AIMultiple, statistical parity with Tavily) | A semantic search API built on its own 500B+ page index, with specialized indexes for people, companies, and academic publications, and page text returned as query-relevant highlights rather than static full-page content. | Exa's structural advantage is specialized retrieval: dedicated indexes for people, companies, and publications produce materially higher top-result recall than Tavily's general web retrieval (75.5% vs. 40.5% on people, 81.5% vs. 61.3% on company, 63.3% vs. 31.8% on academic, per Exa-run benchmarks). Exa returns query-selected highlights rather than static full-page text, which reduces downstream LLM token costs. But Exa's free tier is $10 in credits (about 2,500 Lite searches) with no rollover, not a fixed request count, and its 10 QPS limit on Search is lower than Tavily's published paid-tier RPM. Exa Contents is a separate meter at $1/1k pages, so "content included" applies to search results, not URL extraction. On the independent AIMultiple agentic benchmark, Exa (14.39) and Tavily (13.67) are not statistically distinguishable, so the entity-retrieval gap does not generalize to all query types. | Entity-heavy retrieval (people, companies, academic papers), token-efficient RAG where query-relevant highlights are preferable to static full-page text, and workloads that need specialized indexes rather than general web search. | Free: $10 credits/month (≈2,500 Lite searches), no rollover. Pay-as-you-go: $4/1k Lite or Instant, $7/1k Auto or Fast, $12–15/1k Deep Search, $1/1k Contents pages, $5/1k Answer. Enterprise: custom with ZDR, HIPAA, SSO. |
| Brave Search API Independent agentic benchmark (AIMultiple; Brave only consistent winner over Tavily) Brave Search API pricing (official) | An independent search index with a high-throughput API, built for keyword-oriented retrieval and raw request volume, returning links rather than bundled page content. | Brave runs its own independent index and offers the highest throughput on the entry tier: 50 queries per second on the free plan, with $5 in monthly credit (about 1,000 requests) and a card required for abuse checks. On the independent AIMultiple agentic search benchmark, Brave (14.89) was the only API to consistently outperform Tavily (13.67) by a statistically significant margin—about 1 point—and it also had the lowest mean latency at 669 ms. Brave's documented mode is keyword matching rather than semantic search, and it returns links rather than page text, so the caller must fetch and parse result pages separately. It does not offer a single-call research report comparable to Tavily's Research endpoint. Brave is the throughput and latency choice; Tavily is the bundled-content and unified-credit choice. | High-volume keyword search and low-latency agent loops where throughput and raw request speed matter more than semantic understanding or bundled page content, and where the caller is prepared to fetch and parse result pages separately. | Free: $5 monthly credit (≈1,000 requests), 50 QPS, card required. Paid plans scale by request volume. No per-page content in results; fetching is separate. |
Comparison checked against official product information on 2026-11-04. Features and pricing may change.
Frequently Asked Questions
Is Tavily free?
Tavily has a free Researcher plan with 1,000 credits per month, no credit card required, and monthly resets. That is roughly 1,000 basic searches or 500 advanced searches. Keyless Search and Extract are also available without an account. However, the free plan is a prototyping allowance, not a production tier: Production API keys require a paid plan or Pay as You Go enabled, and the published 1,000 RPM rate limit applies to Production keys, not to the free plan. When credits are exhausted, requests stop until the next monthly reset unless Pay as You Go is enabled.
What does a Tavily Research request cost?
Between 4 and 250 credits, depending on the model and the task. The mini model ranges from 4 to 110 credits per request; the pro model ranges from 15 to 250. The upper bound of 250 credits is a quarter of the free monthly allowance, or 6.25% of the $500 Growth plan's 100,000 credits. Unlike search (1 credit basic, 2 advanced) and Extract (1 credit per 5 URLs), Research does not have a fixed credit price, so the cost of a specific request cannot be computed from the pricing page alone. Teams using Research at volume should instrument per-request credit consumption.
Does Tavily's 1,000 RPM rate limit apply to the free plan?
No. The 1,000 RPM figure applies to Production API keys, and Production keys require an active paid plan or Pay as You Go enabled. The free Researcher plan does not qualify for a Production key. Third-party sources describe the free tier's practical limit as approximately 1 request per second, or about 5 requests per minute, with bursts above the limit returning HTTP 429. Non-paying plans also have a flex tier: search requests that set max_age_hours can be refused with HTTP 434 when Tavily is under heavy load. Do not assume 1,000 RPM is available before upgrading.
How does Tavily perform on people and company search?
According to Exa's published benchmarks (which Exa runs), Tavily's top-result recall is 40.5% on people profiles versus Exa's 75.5%, and 61.3% on company search versus Exa's 81.5%. On academic publication retrieval, Tavily scores 31.8% versus Exa's 63.3%. The gap is attributed to a structural difference: Tavily relies on its general web index for all query types and does not maintain dedicated indexes for people, companies, or publications. This does not mean Tavily is a poor general search API; it means entity-specific retrieval is a weaker use case and should be evaluated on your own queries rather than assumed to be competitive.
Does Tavily train on my queries?
It depends on your settings. Tavily's platform terms permit Tavily and its AI providers to train models on input and output from its AI research features. The 'Allow Use of Query Data' setting controls whether query data is used to improve services: when ON, it may be used; when OFF, it is not stored or used for improvements. Zero Data Retention is a separate capability that must be enabled per account, and contractual ZDR with custom QPS and SLA-backed support is arranged through Enterprise sales. The default posture is not zero-retention, so teams processing sensitive queries must change the setting or negotiate the Enterprise arrangement.
Alternative Tools
-
Exa Exa is a semantic search API built on its own 500B+ page index, with page text included in search results, strong entity-lookup accuracy, an official MCP server, and per-request pricing from $4/1k. It is not a benchmark leader across the board—it is statistically tied with Brave and Firecrawl on independent agentic search benchmarks—and its 10 QPS limit, non-paginating search, crawl-bounded recency, and Enterprise-only compliance features are hard constraints that require workflow design around them. -
Brave Search API Brave Search API provides web, news, image, video, local, and LLM Context results from Brave's independent index, priced per request with a $5 monthly credit, aimed at AI agents and RAG pipelines rather than Google SERP replication. -
Elicit Elicit is a systematic review workflow tool with a strong structured extraction interface, but independent evaluations find its search sensitivity averages 39.5% and its quote/reasoning matching degrades to 10%/0% in high-accuracy mode, positioning it as a complement to human reviewers rather than a replacement. -
NotebookLM NotebookLM (now Gemini Notebook) is Google's source-grounded research assistant: you upload documents, and it answers only from those sources with inline citations, then turns them into Audio Overviews, Video Overviews, Mind Maps, Slide Decks, and study tools. The free tier is genuinely usable; paid limits are bundled inside Google AI plans. The hard boundaries are no consumer API, answers bounded by uploaded sources, and Audio Overview language quality that is still catching up outside English. -
Perplexity Perplexity is a cited answer engine that combines real-time web search with advanced language models, but its $20 Pro tier is licensed for personal non-commercial use only, its Pro Search allowance is an unpublished “weekly limit,” and its API is a separate pay-as-you-go billing system. -
Firecrawl Firecrawl is the most complete LLM-ready web data API: six endpoints covering search, scrape, crawl, map, parse, and interactive browser sessions, with clean Markdown output and first-class MCP support. The advertised 1-credit-per-page rate is a base rate—JSON extraction, PDF parsing, ZDR, and prompt-injection checking multiply it by up to 9x, and Interact bills by the browser minute. Self-hosting is AGPL-3.0 but omits the anti-bot, Agent, and Interact features that define the cloud product. Plan credits do not roll over below $599/month.