Exa: Search API Costs, Source Quality and Alternatives
What does Exa do?
Exa offers search and page-content APIs for developers adding web context to AI applications. Its product lineup also includes research and monitoring capabilities. Start by deciding whether the application needs a ranked set of sources, retrieved page text or a synthesized research result: those are different deliverables with different costs.
Who should consider Exa?
Developers maintaining an assistant that needs references outside its model knowledge, or research teams building a repeatable source-discovery pipeline. An API integration requires application code, credential handling and an output-review process. For a person who simply wants to ask occasional research questions, an interactive research product may require less setup.
How is Exa priced?
As checked on October 26, 2026, the official page lists $10 of monthly free credits plus a $10 onboarding bonus. It lists Lite Search at $4 per 1,000 requests, Auto Search at $7 and Deep Search at $12, with the base price covering up to ten results. Extra results and page summaries can add charges. Treat free credits as a monetary allowance, not an unlimited number of research tasks. Check the live endpoint table before enabling a recurring job.
How can you compare retrieval quality?
Create a small evaluation sheet with ten questions from your real workflow. Include an exact documentation lookup, an ambiguous product name, a recent policy change and a question for which no reliable answer is known. Before making requests, write down acceptable source types and any known reference answers. This prevents judging every plausible-looking result as useful after the fact.
Keep the query, selected endpoint, retrieval date and returned source URLs together. Review source relevance separately from the answer an LLM generates afterward. A relevant source may contain an outdated statement, and a current source may still be summarized incorrectly. For the unanswered question, the desired outcome is an explicit gap rather than a confident invented conclusion.
What should pass before production use?
- Each material answer claim links to a passage that actually supports it.
- Results distinguish product documentation from independent commentary and unrelated namesakes.
- Dates are interpreted correctly: publication date and the date of the described event are not interchangeable.
- Record total API spend and human review time per accepted answer, including retries and rejected results.
- Test how the application behaves when no usable source is returned; do not silently substitute unsupported model knowledge.
What data and cost limits matter?
Queries and any supplied context are sent to an external service. Begin with public or synthetic inputs and review the applicable data terms before including private documents or customer identifiers. Put a request budget around multi-step loops: repeated discovery, content retrieval and summarization can make one user question substantially more expensive than one search request. A source citation is a review aid, not proof that the generated answer is correct.
Which alternatives should you compare?
Compare Tavily for another query-led retrieval API. Evaluate Firecrawl when you already know the URLs and primarily need collection or structured extraction. Consider Perplexity when the intended user wants an interactive research interface. Use the same question set and acceptance rules across candidates; compare complete workflows rather than headline request prices.
Sources checked October 26, 2026: Official product website and official endpoint pricing.