Tavily gives AI agents and LLM apps credit-metered APIs to search the live web, extract and crawl pages, map sites and produce cited research reports, starting with 1,000 free credits a month.
Used this tool? Rate it
Used this tool? Rate it
Tavily is a web search API for AI agents and LLM applications. Instead of handing an agent a list of links, it searches the live web, pulls the relevant page content and returns it in a form a model can use directly. The company describes the product as giving agents secure, at-scale web access, with real-time search, scraping and structured data retrieval in a single API call.
Billing is credit-based across a free monthly allowance, subscription tiers, pay-as-you-go usage and custom enterprise contracts. Access runs through the API, SDKs, a command-line tool or MCP clients.
The Tavily API is organized around five endpoints. Results come back as LLM-ready content rather than raw HTML, and each endpoint draws on the same credit balance at its own rate.
The Research endpoint offers two models. The mini model targets narrow or well-scoped questions, while the pro model handles complex topics that span several subtopics or domains. Developers who need machine-readable results can pass a JSON Schema so the research response is shaped to match it, which makes the output predictable enough to feed straight into another program. Research results can also be streamed as they are generated, as a Server-Sent Events stream.
Tavily claims a 180 ms median (p50) response time on its search endpoint and calls itself the fastest on the market. An independent benchmark that ran its own query set measured Tavily's latency at 998 ms; the two figures come from different measurement setups. On the safety side, Tavily says requests pass through security, privacy and content-validation layers designed to block personal-data leakage, prompt injection and malicious sources.
Automatic parameter tuning can quietly double the price of a search: with it switched on, the search depth may be raised to advanced when that is likely to improve results, which uses 2 credits per request. Setting the depth to basic yourself avoids the extra charge.
Tavily can also report usage per call. When you enable the usage option, the response includes a usage object with credit information, so you can log what each request cost. For pay-as-you-go spending, you can set a monthly usage limit in the billing settings to avoid unexpected costs.
Tavily's published use cases center on agents that need current, cited web information rather than a static database:
These examples come from Tavily's own marketing and documentation. They show what the API is designed for, not independently measured outcomes.
Tavily pricing is built on API credits. You start with 1,000 free API credits per month with no credit card required. Paid subscription tiers then scale from Project at $30 a month to Growth at $500 a month, and the per-credit price falls as volume rises:
| Plan | Credits per month | Monthly price | Price per credit |
|---|---|---|---|
| Researcher | 1,000 | Free | – |
| Project | 4,000 | $30 | $0.0075 |
| Bootstrap | 15,000 | $100 | $0.0067 |
| Startup | 38,000 | $220 | $0.0058 |
| Growth | 100,000 | $500 | $0.005 |
| Pay as you go | Per usage | $0.008 per credit | $0.008 |
| Enterprise | Custom | Custom | Custom |
Pay-as-you-go costs $0.008 per credit and lets you keep being charged per credit once a plan's credit limit is reached.
Search is billed per request, and advanced depth costs twice as much as basic.
| Endpoint | How credits are charged |
|---|---|
| Search | 1 credit at basic depth, 2 credits at advanced depth |
| Extract | 1 credit per 5 successful URLs (basic), 2 per 5 (advanced) |
| Map | 1 credit per 10 successful pages, 2 with instructions |
| Crawl | Mapping cost plus extraction cost |
| Research | 4–110 credits with mini, 15–250 with pro |
Extract counts only successful URLs, billed in blocks of 5. Map requests that fail are never charged, and successful mapping is billed per 10 pages returned. Crawl adds the two together, so crawling 10 pages at basic extraction depth costs 3 credits in total, 1 for mapping and 2 for extraction. Research has the widest spread because its pricing is dynamic within per-request bounds; at the pro model's 250-credit ceiling, a single task uses as many credits as 250 basic searches.
Tavily positions itself against two groups of products. Against traditional search APIs such as Bing, Google or SerpAPI, it says it searches the web, reviews multiple sources and extracts the most relevant content, delivering concise, ready-to-use information for AI applications. Against AI answer engine APIs such as Perplexity Sonar or the OpenAI Web Search API, it stresses customizable search results, developer control over depth, domains and extraction, and LLM-generated answers that are optional rather than the main output.
One independent benchmark compared eight search APIs, retrieving five results per query from each: Brave Search, Tavily, Exa, Firecrawl, SerpAPI, Perplexity and the Base and Pro tiers of Parallel Search. Brave Search led with an Agent Score of 14.89, Firecrawl, Exa and Parallel Search Pro were close enough that the differences could be random variation, and Brave was the only API that consistently outperformed Tavily, by about one point.
Tavily's own pages describe query-data handling in different terms:
Partly. The Researcher plan includes 1,000 credits a month with no card, keyless access lets you try Search and Extract without an account, and students can get four months of the Project plan free.
A basic search uses 1 credit and an advanced search uses 2. Automatic parameter tuning can switch a request to advanced; setting the depth to basic explicitly avoids that charge.
Credits reset on the first day of each calendar month. The pricing, documentation and help pages checked contain no rollover provision.
Nebius announced an agreement to acquire Tavily in February 2026, and a Nebius blog post says Nebius has acquired Tavily. Tavily continues to operate under its own brand.
It depends on your settings and contract. The Privacy Policy allows use of portions of query data unless your contract says otherwise, and the Allow Use of Query Data setting turns this off.
