Direct answer: the bill for AI web search has two parts: a per-request search fee and the context tokens you feed the model. At official prices, 100,000 searches cost about USD 100–1,600 in search fees: Perplexity Fast 100, Exa lite 400, Perplexity standard 500, Tavily basic 500–800, OpenAI Web Search from 1,000, Exa deep 1,200, Tavily advanced 1,600. But if each search brings back 10,000 tokens of page content, model context adds USD 100–10,000 — tokens are usually the bigger line.
This guide is for technical and procurement teams building RAG, agent web access, competitor monitoring or market research. Data cut-off: 1 November 2026; prices come from official Tavily, Exa, Perplexity and OpenAI pages, in USD, excluding tax, with no FX conversion. The examples are editorial models, not quotes; tokens per search, result counts and model prices must be replaced with your own data.
Separate "per request", "per credit" and "per token"; the search fee is usually small — how much page content you pull decides the model bill; batching requests (multiple queries per call) and caching results are the most direct savings.
What does each vendor charge for?
Tavily bills credits (basic search 1, advanced 2); Exa bills per request (lite USD 4/1k, auto 7, deep 12–15); Perplexity per successful request (standard 5, fast 1); OpenAI USD 10/1k plus content tokens.
| Vendor | Billing unit | Official price | Easy to miss |
|---|---|---|---|
| Tavily | Credit | PAYG USD 0.008/credit; plans down to 0.005; basic search 1 credit, advanced 2 | Research uses 4–250 credits per request |
| Exa | Request (first 10 results included) | lite USD 4/1k; auto/fast 7; deep 12–15 | Results above 10 add USD 1/1k; summaries USD 1/1k pages |
| Perplexity | Successful request (up to 5 queries = 1 unit) | Standard USD 5/1k; Fast 1/1k | Batch queries dilute cost; empty results still bill |
| OpenAI | Call count + content tokens | Web Search USD 10/1k plus content tokens | Tokens billed at the model's rates |
Sources: Tavily pricing and credits docs, Exa pricing, Perplexity pricing, OpenAI API pricing (1 November 2026). Public list prices, excluding tax and enterprise agreements; each vendor defines a "request" differently, so check before signing.
What do 100,000 searches cost?
Search fees alone: Perplexity Fast about 100, Exa lite 400, Perplexity standard 500, Tavily basic 500–800, OpenAI from 1,000, Exa deep 1,200, Tavily advanced 1,600.
| Option | Unit price | 100,000 searches | Note |
|---|---|---|---|
| Perplexity Fast Search | USD 1/1k | 100 | Lower latency, same response shape |
| Exa lite | USD 4/1k | 400 | First 10 results included |
| Perplexity standard | USD 5/1k | 500 | Up to 5 queries per request = 1 unit |
| Tavily basic (Growth tier) | USD 0.005/credit | 500 | 1 credit per request |
| Exa auto/fast | USD 7/1k | 700 | Automatically picks strategy |
| Tavily basic (PAYG) | USD 0.008/credit | 800 | Pay as you go |
| OpenAI Web Search | USD 10/1k | From 1,000 | Plus content tokens |
| Exa deep | USD 12/1k | 1,200 | Multi-step research, structured output |
| Tavily advanced (PAYG) | USD 0.016/request | 1,600 | 2 credits per request |
Editorial model: unit price × 100,000 ÷ 1,000; excludes tax, retries, parsing and model tokens. If Perplexity's "up to 5 queries per request" is used, the effective unit price drops further.
Why do tokens usually cost more than the search fee?
Because results are fed to the model as context. At 10,000 tokens per search, 100,000 searches is 1 billion tokens: USD 100 at 0.10/M, 2,000 at 2/M, and 10,000 at 10/M — easily above the search fee.
| Context per search | Total tokens (100k searches) | At USD 0.10/M | At 2/M | At 10/M |
|---|---|---|---|---|
| 2,000 tokens | 200M | 20 | 400 | 2,000 |
| 5,000 tokens | 500M | 50 | 1,000 | 5,000 |
| 10,000 tokens | 1B | 100 | 2,000 | 10,000 |
| 20,000 tokens | 2B | 200 | 4,000 | 20,000 |
Editorial model: tokens = context per search × 100,000; model prices are example tiers, not a vendor quote. Actual billing follows the model invoice; caching, truncation and summaries can cut input sharply.
How are content and result counts charged?
Exa adds USD 1/1k for results beyond 10, USD 1/1k pages for summaries, and USD 1/1k pages per content type for Contents; Tavily charges 1 credit per 5 successful extractions and 1 credit per 10 mapped pages; Perplexity's extracted text length does not change the request fee.
| Operation | Official rule | Conversion |
|---|---|---|
| Exa extra results | USD 1/1k beyond 10 results | 20 results = base + 10 extra |
| Exa page summaries | USD 1/1k pages | Each summary billed separately |
| Exa Contents | USD 1/1k pages per content type | text + highlights = two types |
| Tavily Extract | 1 credit per 5 successful extractions (basic) | About USD 0.0016/URL (PAYG) |
| Tavily Map | 1 credit per 10 pages | Map before crawling to save credits |
| Perplexity text length | Does not change the request fee | But it raises your model tokens |
Sources: official pricing and docs; Tavily advanced extraction is 2 credits per 5. The table only converts units, not a recommendation.
Which scenario fits which option?
For links and snippets: Exa lite or Perplexity Fast. For batched multi-query: Perplexity (5 queries per request). For multi-step research: Exa deep or Tavily Research. For controlling generation: a search API plus your own model.
| Scenario | Start with | Why | Do not ignore |
|---|---|---|---|
| Agent fetching links fast | Perplexity Fast / Exa lite | USD 0.001–0.004 per call | Snippet length raises tokens |
| Batched multi-query retrieval | Perplexity (5 queries/request) | Dilutes one request across 5 queries | Empty results still bill |
| Multi-step deep research | Exa deep / Tavily Research | Structured output and citations | USD 12–15/1k or 4–250 credits |
| Full page text only | Exa Contents / Tavily Extract | Priced per page or URL | Content types bill separately |
| Own generation pipeline | Search API + your model | Generation cost stays controllable | Tokens are the main line |
Editorial guidance, not a vendor ranking; compare result quality, latency and total cost per query (search fee + tokens) on the same query set.
What should procurement and acceptance check?
Billing unit, queries per request, whether results and content are extra, failure billing, caching and rate limits — not just "price per call".
| Acceptance item | Suggested definition | Evidence to keep |
|---|---|---|
| Billing unit | Per call/credit/token; queries per request | Price list, sample invoice |
| Results and content | First N results included; summaries and full text extra? | Request params, invoice |
| Failure billing | Empty results, 429s, upstream failures | Official terms |
| Context size | Tokens returned per call; can it be capped | Response sample, model invoice |
| Caching | Can identical queries be cached, for how long | Architecture doc |
| Compliance | Source domains, copyright, personal data | Terms, review records |
| Latency | P95, concurrency, rate limits | Load test |
| Exit | Result format, migration cost | Export sample |
Editorial guidance, not legal advice; crawling and quoting third-party content must follow source-site terms and applicable law.
Who should read this — and who should not?
For teams building RAG, agent web access and market research. Not for static scraping only, or projects without a query volume and result-length target.
Read if: you are in engineering, data or procurement; you are comparing Tavily, Exa, Perplexity and OpenAI Web Search; you need a budget for 100,000 searches. Skip if: you only scrape fixed sites (a collection API is cheaper); you have no query sample; you want this article to pick a vendor. Supports: separating search fees from token fees, batching and caching strategy, and acceptance criteria.
Lingyi does not rank search APIs or promise result quality. To connect retrieval, permissions and model routing to an enterprise knowledge base or agent, see LingHarness; to fold search cost into project budgets, talk to us.
Related reading: vector search cost, data collection cost, and RAG knowledge-base total cost; this article only covers web-search API pricing and token sensitivity.
FAQ
How is AI web search billed?
Three models: Tavily per credit (basic 1, advanced 2, PAYG USD 0.008/credit); Exa per request (lite 4, auto 7, deep 12–15 per 1,000); Perplexity per successful request (standard 5, Fast 1 per 1,000, up to 5 queries per request = 1 unit); OpenAI USD 10/1,000 plus content tokens.
What do 100,000 searches cost?
Search fees alone are about USD 100–1,600: Perplexity Fast 100, Exa lite 400, Perplexity standard 500, Tavily basic 500–800, OpenAI from 1,000, Exa deep 1,200, Tavily advanced 1,600. Add model context tokens on top.
Why do tokens cost more than the search fee?
Because results are fed to the model. At 10,000 tokens per search, 100,000 searches is 1 billion tokens: USD 2,000 at 2/M and 10,000 at 10/M, far above the USD 100–1,600 search fee. Capping returned length and caching beats switching vendors.
How should Perplexity's "5 queries per request" be used?
One POST can carry up to five queries and is billed as one unit. Batching queries on the same topic dilutes the effective unit price to one fifth; but empty results still bill, so validate the queries.
Which is cheapest for full page text?
Exa Contents at USD 1/1,000 pages per content type; Tavily Extract at 1 credit per 5 successful extractions (about USD 0.0016/URL PAYG). For fixed sites, a dedicated collection API priced per page may be cheaper.
Are failed requests billed?
It varies: Perplexity states that 429s, invalid requests and upstream failures are not billed, but a successful response with no results is; Tavily does not charge failed extractions. Confirm empty results, timeouts and retries before signing.
How should we estimate our own bill?
Two lines: search fee (requests × price) + token fee (context per search × volume × model price). Run 500–1,000 real queries to measure average result length, then extrapolate — not just "price per call".
Method and limitations
This article uses only official pages accessible on 1 November 2026: Tavily pricing and credits documentation, Exa pricing, Perplexity pricing and OpenAI API pricing. All amounts are converted from public list prices and example assumptions; no enterprise agreements or reseller quotes are used. USD is not converted to RMB and no vendor ranking is implied. Each vendor defines "request", "result" and "content type" differently; actual billing follows the official docs and invoice. Tokens per search, result counts and model prices are examples, not benchmarks; caching, truncation and summaries change real input. The article does not assess search quality, freshness or compliance and is not legal advice.
Sources and basis
Data cut-off: 1 November 2026. The following are official pages; prices, units and terms follow the original text and invoice.
- Tavily, Pricing and Credits: tavily.com/pricing, docs.tavily.com/documentation/api-credits. PAYG USD 0.008/credit, plans 0.005–0.0075; basic search 1 credit, advanced 2; extraction 1 credit per 5; Research 4–250 credits.
- Exa, Pricing: exa.ai/pricing. lite 4, auto/fast 7, deep 12–15 per 1,000 requests; Contents USD 1/1,000 pages per type; extra results USD 1/1,000; summaries USD 1/1,000 pages.
- Perplexity, Pricing: docs.perplexity.ai/getting-started/pricing. Search API standard 5, Fast 1 per 1,000; up to 5 queries per request = 1 unit; 429s and failures not billed.
- OpenAI, API pricing: platform.openai.com/docs/pricing. Web Search USD 10 per 1,000 calls plus content tokens.