Cheapest search APIs: compare the bill, not the headline rate
All Them APIs
Independent API research
Several search APIs publish a $1-per-thousand rate. They are not necessarily the same product, the same request or the same amount due today.
Parallel Fast and Turbo and Perplexity Fast Search list that request rate. Serper's Starter pack also works out to $1 per thousand, but requires buying a larger pack. Context.dev meters results and content options through plan credits. Tavily charges different credits for basic and advanced search.
The useful answer to "which is cheapest?" starts with your workload and ends with a bill. The search vendor directory separates agent retrieval from SERP data services; do not assume the lower number buys the same output.
Published prices, with the unit left attached
USD documentation reviewed October 27, 2026. These are list rates and explicit arithmetic, not observed charges from a new benchmark. Free allowances, tax, infrastructure and separate model usage are not included in the request-rate column.
| Provider/mode | Published rate or effective allowance | What changes the comparison |
|---|---|---|
| Parallel Turbo / Fast | $1 / 1,000 requests | Default ten results; additional results and separate tools add usage |
| Perplexity Fast Search | $1 / 1,000 successful requests | Different mode from standard Search; successful empty results count |
| Exa Lite / Instant | $4 / 1,000 requests | Up to ten results; extra results and some content options add cost |
| Perplexity standard Search | $5 / 1,000 successful requests | Raw sources, not a generated answer |
| Brave Search | $5 / 1,000 requests | Search allowance and selected plan apply |
| Exa Auto / Fast | $7 / 1,000 requests | Different retrieval configuration from Lite / Instant |
| Tavily basic, PAYG | $8 / 1,000 requests | One credit each at $0.008 per credit |
| Tavily advanced, PAYG | $16 / 1,000 requests | Two credits each at the same PAYG credit rate |
| Context.dev Developer | $20/month includes 10,000 credits | Search, highlights and Markdown consume different units |
| Serper Starter | $50 / 50,000-query pack | $1 / 1,000 only at full pack use; credits expire |
| SerpApi Starter | $25/month / 1,000 searches | Monthly quota, feature coverage and throughput differ |
Sources: Parallel, Perplexity, Exa, Brave, Tavily, Context.dev, Serper and SerpApi.
Minimum commitment can dominate a small workload
Suppose your application makes 1,000 successful requests and then you decide not to continue.
A published $1-per-thousand usage rate implies $1 of metered usage, but it does not prove the provider lets you fund exactly $1. Check minimum credit purchases and billing setup separately. A $50 prepaid pack still means paying $50 even if the amount you actually consume corresponds to $1 of its advertised capacity.
Likewise, a monthly plan can have an attractive fully utilized unit cost and be expensive for an intermittent job. Divide the amount you actually pay by the useful requests you actually make. Do not divide by an allowance that expires unused.
This is why our free-search guide treats trial allowances separately from sustainable paid usage.
A credit is not a shared unit across vendors
Tavily basic and advanced consume different amounts of the same balance. Context.dev's Search, highlights and Markdown options have different credit costs. The number printed in a dashboard cannot be compared without the operation it funds.
For example, the Context Search product says a ten-result search uses one credit; adding highlights adds one, and reading Markdown can add one per result page. A source-only lookup and a ten-page reading job are not two examples of the same one-credit operation.
Ask the agent to record returned usage fields where available. Keep unknown cost unknown rather than calculating an exact number from an account balance shared with other work.
Lower request price does not establish lower task cost
Imagine configuration A uses two calls to finish a task and B uses ten. The correct comparison includes all calls, extraction and model tokens, with failed tasks still counted.
Do not turn that example into an assumed provider difference. It is a reason to measure the application. A richer response can save work, but it can also add irrelevant context and increase model spend.
Our benchmark register keeps request pricing separate from external dollars-per-task metrics. Vendor benchmark cost columns can include an answer model, repeated searches or a high-volume subscription rate. Copying that number into a "per API call" table changes its meaning.
The lowest published rate is not always available on your plan
Some benchmark tables use enterprise or bulk tiers. Others divide an annual subscription by the full annual allowance. Some include free calls. Check commitment, expiry, rate limits and whether content extraction is included before treating those figures as a price you can buy today.
Context.dev's older external benchmark entries illustrate the problem: a benchmark's recorded plan economics need not match the current Developer or Pro page. We use today's provider documentation for pricing and leave old rates attached to their dated benchmark record.
A practical low-cost shortlist
For raw agent retrieval, evaluate Parallel Fast/Turbo and Perplexity Fast at their documented rates, then compare source quality with a low-cost Exa mode. For a Google SERP workflow, compare the required output and pack economics separately; a general source list is not a Google-rank product.
For a broader search-and-reading workflow, compare the full credits consumed by Context.dev, Tavily, Exa Contents or another extraction path. Add concurrency and throughput requirements before choosing a plan.
The next decision is concrete: what is the smallest real task you can complete on each provider, and what did completing it actually consume? Our pair comparisons help select the first candidates without inventing a universal price winner.