Tavily vs Exa vs Perplexity Sonar API: Real AI Web Search Cost per 100,000 Queries in 2026
The short answer: Among the three products compared here, Tavily and Exa both price simple search retrieval in the same narrow band, about $0.007–$0.008 per query, and Perplexity's flat Search API undercuts both at $0.005. On October 2, 2026, Cloudflare added a fourth route to the market: a Web Search API inside AI Gateway, in open beta, that sends queries to Ceramic.ai, Linkup or Exa at each provider's list price with no per-search markup. Ceramic.ai, the default provider, lists at $0.25 per 1,000 requests, which is 20 times below Perplexity's flat Search API, 28 times below Exa and 32 times below Tavily's pay-as-you-go credit.
This comparison covers search and retrieval services for your pipeline. For model-generated answers with web sources, compare grounded-answer billing. OpenAI Web Search vs Gemini Grounding vs Perplexity Sonar: Real Grounded AI Search Cost per 100,000 Queries in 2026.
That spread describes the billed retrieval call only. It says nothing about result quality, depth, latency, payload size or index coverage, and the cost of passing results to a language model can outweigh the retrieval fee entirely. The multiplier that separates the original three products is still deep content extraction and, especially, Tavily's research endpoint at 4 to 250 credits per call, a 62.5x range inside one product.
Perplexity's Sonar API returns a synthesized answer instead of raw results, so it adds a per-request fee on top of token charges, and that fee varies 2.4x by context size. None of these figures include a separate downstream LLM call to generate a final answer from raw search results, except where the product (Sonar) does that synthesis itself.
My practical rule of thumb: don't pick a search API on the headline per-query price alone. A cheap call that returns mushy results, and makes your LLM read twice as much text, can easily cost more than a pricier call that gets it right the first time.
What each vendor bills
Tavily. Pay-as-you-go at $0.008 per credit. Basic search costs 1 credit, and advanced search (deeper crawling) costs 2 credits. Extract is 1 credit per 5 URLs, and Map is 1–2 credits per 10 pages. The research endpoint runs 4–110 credits per request. Subscription plans lower the per-credit rate at volume:
- Project: $30/mo, 4,000 credits, $0.0075/credit
- Bootstrap: $100/mo, 15,000 credits, $0.0067
- Startup: $220/mo, 38,000 credits, $0.0058
- Growth: $500/mo, 100,000 credits, $0.005
Exa. Fixed per-endpoint rates: $7 per 1,000 search requests ($0.007 each) and $5 per 1,000 answer requests ($0.005 each). New accounts receive $20 in starting credits plus $10 of free credits monthly. Requests fail closed when the balance runs out, so you'll see an error rather than a surprise bill.
Perplexity Sonar API. This one is a hybrid meter: $1 per million input tokens and $1 per million output tokens, plus a per-request fee of $5 to $12 per 1,000 requests depending on context size ($0.005–$0.012 per request). A separate flat Search API without synthesis costs $5.00 per 1,000 requests ($0.005 each). Sonar returns pre-synthesized answers with citations, so LLM generation is bundled directly into the search call.
Cloudflare Web Search API: one gateway, three providers
Cloudflare's Web Search API routes queries through AI Gateway to third-party providers with zero per-search markup. You get one integration point and a choice of backends, which is handy if you'd like to switch providers without rewriting your agent code.
What Cloudflare provides, and what the provider provides
| Provider | List price per 1,000 requests | Per request | How it is called through Cloudflare | What each result contains | Zero Data Retention |
|---|---|---|---|---|---|
| Ceramic.ai (default) | $0.25 | $0.00025 | Ceramic's own index of more than 40 billion pages | Page descriptions up to 8,000 characters | Yes |
| Linkup | $5.00 | $0.005 | fast search depth, raw search results with no generated answer | Sourced snippets | Yes |
| Exa | $7.00 | $0.007 | auto search type | Page highlights returned as the description | No |
Cost Comparison at 100,000 Queries
| Search Route | Base Rate per 1k Queries | 100,000 Queries Cost | Output Format | Best Suited For |
|---|---|---|---|---|
| Cloudflare Ceramic.ai | $0.25 | $25.00 | Raw metadata snippets | Large-scale pre-filtering |
| Perplexity Search API | $5.00 | $500.00 | Clean snippets | Budget RAG pipelines |
| Exa Search API | $7.00 | $700.00 | Neural semantic highlights | Financial and code search |
| Tavily Basic Search | $8.00 | $800.00 | Pre-parsed clean markdown | Autonomous agent tool calls |
| Perplexity Sonar Synthesized | $5.00 + $1/M tokens | ~$800 – $1,200 | Direct answer with citations | End-to-end Q&A assistants |
Interactive Matchup Simulator
If you'd rather not do this math on a napkin, you can model your app's exact monthly search and RAG synthesis budget across different query volumes with our live tool:
- Test your token volumes and search frequencies on the Interactive Search API Matchup Simulator to compare Tavily, Exa, Perplexity Sonar, and Google Vertex Search side by side.
Engineering Recommendations
- For Agentic Workflows (LangGraph, CrewAI): Choose Tavily. Its pre-cleaned markdown outputs fit cleanly into prompt context windows without custom HTML scraping.
- For High-Precision Semantic Search: Choose Exa. Its neural embedding index excels at finding technical repositories and niche corporate filings.
- For Direct Q&A Chatbots: Choose Perplexity Sonar. Eliminating downstream LLM calls saves 60% to 70% in end-to-end system latency and token spend.
If I had to boil it down: start with whichever tool matches the shape of your workload (agent tool calls, semantic lookup, or finished answers), then test the cheaper routes on a sample of your real queries. If the quality holds up, take the savings. If it doesn't, the extra few hundred dollars per 100,000 queries is usually cheaper than debugging bad retrieval later.