What does Tavily return?
Search returns sources and content with optional answer/raw-content settings. Extract, Crawl, Map and Research are different tasks rather than synonyms for a Search request.
Tavily is a sensible candidate when you want a single provider for discovering sources and extracting selected URLs. Keep the first test small: a source list that your agent can inspect, followed by extraction only where it needs more context.
Pricing and trial limits
- Basic Search costs one credit; advanced costs two. Pay-as-you-go is $0.008 per credit.
- That is $8 per 1,000 basic requests or $16 per 1,000 advanced requests at the published PAYG credit rate, before other operations.
- The free allowance is 1,000 credits per month without a card. Credits are not necessarily 1,000 advanced searches.
These are documented prices, not spend measured by our own run. Check the current pricing source and the exact request settings before comparing a request with a credit or a completed answer.
Integration decisions for a coding agent
- Set search_depth and the optional answer/raw-content fields deliberately. An application requesting only sources should not rely on answer-generation defaults.
- The official Python and JavaScript clients are convenient entry points; a direct server-side request is also documented. Keep credentials in the runtime environment.
- If automatic parameter selection is enabled, inspect the actual mode and credit usage. A fixed query count alone will not explain the bill when depth can change.
For an assistant connection rather than an application backend, read our web-search MCP guide. MCP availability does not demonstrate the quality of a particular agent/client integration.
What benchmark coverage tells us
No All Them APIs head-to-head retrieval score is assigned to Tavily yet. Public benchmark adapters choose different depth, content and model settings, so a low score from one harness is not a provider-wide verdict. Test the exact configuration you would ship.
The benchmark-source register identifies publishers, scope and known issues. Only Exa, Perplexity and Parallel have our own nine-attempt retrieval reports. Documentation review is not a new performance test.
What to test before committing
- Run basic and advanced against the same recognizable sources.
- Count credits and inspect whether the extra depth changed useful evidence.
- Keep empty results, timeout and authentication failure as distinct outcomes.
- Then add the agent and measure whether optional extraction reduces or increases follow-up calls.
Save enough context to explain a failure: query, mode, result cap, requested location, response status, duration and usage. Exclude keys and private user inputs from public logs. Repeating a fixed request is more informative than quietly changing parameters until it passes.
Tavily alternatives
- Brave for web results and a separate LLM Context surface
- Exa for multiple retrieval modes and Contents
- Context.dev for optional highlights and Markdown within Search
Choose an alternative because its output, cost or integration contract fits the job better. A broad brand comparison can hide that you switched from a source-retrieval API to a generated-answer product.
Sources and review dates
Documentation review: October 27, 2026. Our own measured retrieval runs: August 23, 2026. External reports retain their publisher and measurement window.
Corrections: email faizank@docsalot.dev with an official source.