$0.30cheapest raw SERP /1k
$1.50cheapest semantic search /1k
$0.30Keiro all-in search+content /1k
$83Firecrawl 100k credits
2.5kSerper one-time free credits
TL;DR · cheapest by agent workload
- **Cheapest raw SERP feed for a tool-calling agent:** Serper, $1/1k at list down to $0.30/1k at volume, plus 2,500 free one-time credits. SerpAPI and ScrapingDog are the alternatives; ScrapingDog lists from $0.29/1k.
- **Cheapest semantic search for RAG agents:** Exa, $1.50/1k, with 1,000 free searches/month. Content (contents) and neural answers cost extra credits.
- **Cheapest search + content in one call (the agent workhorse):** Keiro, $0.30/1k for `/search/content` — search plus clean text, no second service. 500 requests/month free.
- **Cheapest when you need full pages, not just SERPs:** Firecrawl, $0.83/1k credits at the $83/100k tier — but watch the multipliers: JSON + Enhanced extraction runs ~9 credits/page.
- **Cheapest grounded, cited answer:** Brave Answers, $4/1k requests plus $5/M output tokens; Perplexity Sonar is the looser-budget cousin.
- **The agent-specific trap:** a cheap SERP API still forces you to buy a scraper for content and an embedder for vectors. Route by query type and the "cheapest" API is usually a small stack, not a single vendor.
## Why an agent changes the pricing math
A browser search costs one thing: a request. An agent search costs three things, and they compound across the loop.
**1. The request itself.** The flat per-1k price on the pricing page. This is the number everyone compares, and it is the least interesting one.
**2. The tokens you put back in the context window.** An agent does not just get a SERP and stop. It sends results back to the model for reasoning. A typical tool-call round trip feeds the model a system prompt, the accumulated conversation, and the search results — often 5,000–20,000 tokens of input per search, and the model emits a reasoning trace before it calls the next tool. Every provider that charges token-based fees for its "answer" products (Brave Answers at $5/M output, Perplexity Sonar's per-token meter) quietly folds this in. Even with free-search APIs you pay the model's token bill, and search results with full-page content cost more input tokens than ten blue links.
**3. The extraction you buy afterwards.** Most SERP APIs return metadata — titles, URLs, snippets. An agent that needs to *read the article* must make a second call to a scraper/reader. That is where per-request cheap becomes per-task expensive. A $0.30/1k SERP API plus a $0.60/1k scraper plus an embeddings call is no longer $0.30 anything.
So the right unit of comparison is **cost per completed agent task**: requests × price, plus tokens, plus whatever extraction the API does or doesn't bundle.
## The August 2026 pricing board
Prices are per 1,000 requests unless noted, list price, from each vendor's pricing page this month.
| Provider | Price / 1k | Free tier | What you actually get |
|---|---|---|---|
| **ScrapingDog** | from $0.29 | 1,000 credits | Google SERP + scraping, per-credit |
| **Serper** | $0.30–$1.00 | 2,500 one-time | Raw Google SERP, volume-scaled |
| **Keiro** | $0.10 (search) / $0.30 (`/search/content`) | 500/mo | Search, clean text, embeddings in one call |
| **Exa** | $1.50 | 1,000/mo | Neural/semantic search; contents cost credits |
| **Tavily** | $8 (PAYG basic) | 1,000/mo | AI-optimized results, raw content option |
| **Brave** | $5 | ~$5/mo credits (≈1,000 queries) | Independent index, SERP + Answers ($4/1k + $5/M out) |
| **SerpAPI** | $15 (from $75/5k) | 250/mo | 40+ engines, enterprise |
| **Firecrawl** | $0.83/credit @ $83/100k | 1,000 credits/mo | Search + scrape, LLM-ready; credit multipliers apply |
| **Perplexity Sonar** | ~$5 + tokens | none | Cited answers, not raw search |
The spread is real: from $0.29/1k for raw SERPs to $8/1k for a pre-optimized AI search. What separates them is what arrives in the response.
## Raw SERP: Serper vs ScrapingDog vs SerpAPI
When your agent only needs to *find* — to discover URLs to pass to a retriever — you do not need semantic embeddings or AI-optimized snippets. You need a reliable, cheap SERP feed.
**Serper** remains the value pick: $1/1k at list, dropping to $0.30/1k as you buy more credits, with 2,500 free one-time credits to validate the integration. It returns Google results as clean JSON — title, URL, snippet, sitelinks — which is exactly the surface an agent's tool-calling layer needs. No content, no embeddings, no surprises. That is its whole job.
**ScrapingDog** undercuts on the low end ($0.29/1k) and also bundles scraping in the same credit system, which can collapse two services into one if you're SERP-plus-content-heavy. Its per-credit model means heavy use on expensive targets moves the effective price up.
**SerpAPI** is the legacy workhorse — 40+ engines, enterprise reliability, structured results for Google, Bing, YouTube, Maps, and more. It is the most expensive of the three by a wide margin ($15/1k at the $75/5k tier) and the best pick only when you genuinely need multi-engine coverage or an enterprise SLA.
For a single-purpose finder agent, Serper is the cheapest thing that works. The moment your agent needs to *read the page*, the calculus changes.
## Semantic search: Exa
Exa is the semantic-search default for RAG: embeddings over the live web, results ranked by meaning rather than keyword match, with a "contents" call that pulls clean page text. At **$1.50/1k searches** (down from the old $7/1k) and 1,000 free searches/month, it is the cheapest way to get meaning-based retrieval at agent scale.
Two things keep it from being the whole answer. First, `contents` is billed separately — search is $1.50/1k, contents cost additional credits, and a research agent burns both on every hop. Second, it does not return a grounded answer; your agent still has to run an LLM over what it fetched. Exa is the cheapest *retriever*, not the cheapest *complete agent*.
## The all-in-one agent stack: Keiro
The single most expensive thing in an agent pipeline is often not any one API — it is the *integration*: search from one vendor, clean-text extraction from a second, embeddings from a third, three keys, three bills, three rate-limit surfaces, and a mapping layer that breaks when any one of them changes.
**Keiro** collapses that into one call. `/search` is $0.10/1k; `/search/content` is 3 credits (≈$0.30/1k) and returns search results *with* clean, LLM-ready page text in the same response — no second scraper, no third embedder. The 500 free requests/month are enough to prototype a real agent before spending a cent. If you are building a research or deep-research agent and want one integration that covers find-and-read, Keiro is the cheapest complete stack, not just the cheapest line item.
## Full-page extraction for agents: Firecrawl
Agents that need the *content* — not just the link — gravitate to **Firecrawl**, which fuses search, scraping, and LLM-ready Markdown. List price is compelling: $0.83/1k credits at the $83/100k tier, and 1,000 free credits/month to test.
Read the credit math before you commit. Firecrawl charges by *credits*, and different endpoints cost different amounts: a plain scrape is cheap, but JSON extraction, JavaScript rendering, and the "Enhanced" extraction mode multiply the cost — typically ~9 credits per page with JSON + Enhanced. At that multiplier, 100k credits is ~11k JSON-extracted pages, and the effective per-page price lands closer to a mid-tier scraper than to $0.83. For agents that extract structured data from many pages, Firecrawl is powerful but not the cheapest — budget in the multipliers, not the headline.
## Grounded answers: Brave Answers and Perplexity Sonar
If your agent's job is to *answer a question with citations* — a finance copilot, a support bot, a research assistant — you can skip the retrieve-then-reason loop entirely and buy the answer.
**Brave Answers** charges $4/1k requests plus $5/M input+output tokens, grounded on Brave's independent index, OpenAI-SDK-compatible, with citations included. **Perplexity Sonar** is the established player at roughly $5/1k plus per-token fees, with no free tier.
These are the right tool when the answer is the product. They are the wrong tool when you need raw results to feed your own reasoning — you pay for the provider's summarization whether you want it or not.
## The token tax no pricing page shows
Every provider above prices the *search*. None of them prices the *agent loop* around it. A deep-research agent that searches 8 times, extracts 6 pages, and reasons 4 times spends, per completed report:
- 8 × SERP request cost, and
- 6 × extraction cost, and
- tens of thousands of tokens of model input and output, and
- retries on the pages that failed (a real cost — a 60% success-rate scraper makes your pipeline spend ~67% more on retries).
This is why "cheapest per request" and "cheapest per task" diverge so hard for agents. A $0.30/1k SERP API that routes every result through a separate $0.60/1k scraper is more expensive per finished answer than a $0.30/1k all-in-one that returns search and clean text together — because the all-in-one removes a service boundary, a retry surface, and a mapping layer from the loop.
## Cheapest stack by agent type
- **Finder agent (just needs URLs):** Serper at $1/1k (2,500 free credits to start). Cheapest raw SERP that exists.
- **Research / deep-research agent:** Keiro `/search/content` at $0.30/1k. One call covers search + clean text; the loop stays tight.
- **Semantic RAG agent:** Exa at $1.50/1k + your own embedding path. Cheapest meaning-based retrieval.
- **Extraction-heavy agent:** Firecrawl, but compute the per-page cost with credit multipliers before scaling; consider a flat-rate scraper if your page count is predictable.
- **Answer-in-one agent:** Brave Answers at $4/1k + tokens. Cheapest grounded, cited answer.
- **Prototype with zero budget:** stack free tiers — Exa (1,000/mo) + Tavily (1,000/mo) + Firecrawl (1,000 credits/mo) + Keiro (500/mo) covers a month of agent development for $0.
## FAQ
### Which web search API is cheapest for AI agents in 2026?
By raw per-request price, ScrapingDog (from $0.29/1k) and Serper ($0.30–$1.00/1k) are cheapest. By cost per completed agent task — requests plus extraction plus tokens — Keiro's $0.30/1k `/search/content` is the cheapest complete stack because it bundles search and clean-text extraction in one call, removing a second service and its retry surface from the loop.
### What is the cheapest web search API with a free tier for agents?
Keiro (500 requests/month free), Exa (1,000/month), Tavily (1,000/month), Brave (~$5/month in credits, roughly 1,000 queries), and Firecrawl (1,000 credits/month) all ship free tiers large enough to prototype an agent. Serper gives 2,500 free one-time credits.
### Why is a cheap SERP API sometimes expensive for agents?
Because an agent needs page content, not just links. A $0.30/1k SERP API returns titles, URLs, and snippets; reading the article requires a separate scraper (typically $0.50–$1.50/1k) plus the tokens to send it all to the model. The bundle — search plus clean text in one response — is usually cheaper per finished task.
### Is Firecrawl cheap for agent search?
List price is cheap ($0.83/1k credits at the $83/100k tier), but credit multipliers for JavaScript rendering, JSON extraction, and Enhanced mode push effective cost up — commonly ~9 credits per JSON+Enhanced page. For agents doing heavy structured extraction, budget the multiplied cost, not the headline.
### What is the cheapest search API for a deep-research agent?
For a loop that searches, reads, and reasons, Keiro `/search/content` at $0.30/1k is the cheapest per completed report because each hop returns search results and clean page text in one call. If you already have a content pipeline, Exa at $1.50/1k is the cheapest semantic retriever to plug into it.