Claude and OpenAI web search both cost $10 per 1,000. Serper’s entry pack is $1.

Claude’s API and OpenAI’s current web_search tool both list $10 per 1,000 searches. You then pay for the text those searches put into the context window. Serper’s entry pack is $1 per 1,000 Google result-page queries. If the harness only needs that result page, the native tool fee is ten times Serper’s, before a single token.

OpenAI’s older web_search_preview tool, on models that are not reasoning models, is $25 per 1,000, and on that path the search-content tokens are free. The web-search guide says new Responses integrations should send { "type": "web_search" }, which is the $10 path.

The consumer products, ChatGPT search and Claude web search inside the apps, are a different bill. That comparison is Best AI agent for web search. This page is the API a builder wires into an agent loop. Every figure below was read from the vendor’s pricing or docs page on November 11, 2026. A per-1,000 figure that divides a plan price by the credits in the plan is this page’s arithmetic, and it assumes the plan is used up.

Which tool do you actually call?

Match the object you get back. A snippet list, a cached page, a Google result page, and a cited answer are four different tool results. They do not become the same thing because each one is called search.

The $10 native fee is the right buy when you do not want to host the tool loop. It is the wrong buy when the next step only needs titles and URLs. Tool use is the model asking for a result. You still choose which result, and which bill, that call returns.

What does each 1,000 searches cost on November 11?

The table is the list price of one successful search call, not a finished agent task. Native tools add tokens. Credit plans get cheaper, or more expensive, when you do not use the pack.

Serper’s other packs, from the same homepage, are Standard at $375 for 500,000 credits ($0.75 per 1,000, 100 queries a second) and Scale at $1,250 for 2.5 million ($0.50 per 1,000, 200 a second). Prices exclude sales tax. Credits are deducted on a successful response, and a zero balance stops new queries.

SerpAPI’s pricing page continues through Cloud 1M at $3,750 a month for 1,000,000 searches, which is $3.75 per 1,000 if that plan is used up, at 110,000 searches an hour. Ludicrous Speed is a 2× multiplier on the plan price. The $25 and $10 figures above are Best Effort, the default.

What do you get back, besides a URL?

This is the cut that the sticker price hides.

Tavily’s search reference, read the same day, says basic and fast return chunks of about 500 characters, advanced about 800, and chunks_per_source stops at 3. Ultra-fast returns one NLP summary per URL. max_results stops at 20. Full-page extraction is the Extract endpoint, and this table does not price it. Basic, fast, and ultra-fast cost 1 credit. Advanced costs 2.

Exa’s pricing page prices /search with lite at $4 per 1,000 requests for up to 10 results, and it does not add an AI-summary charge on that row. The same docs say lite serves page content from the cache and never fetches a page live. auto and the legacy fast type are $7 per 1,000, and an AI page summary on those types is $1 per 1,000 pages. A result past the tenth is $1 per 1,000 results on both.

Brave’s Search plan says the $5 buys complete search results, named as URLs, text, news, images, and more, plus LLM context, at 50 queries a second. The help page says each plan includes $5 in free credits every month, and that the Search credit is 1,000 queries. It also says there is no standalone free plan: the card is how the $5 credit is turned on, with the prepay amount set to $0. Answers is a different plan, $4 per 1,000 requests plus $5 per million input and output tokens, at 2 queries a second. This comparison prices Search.

Perplexity’s pricing page, read November 11, does not list a Sonar Chat Completions request fee. What it lists is a Search API at $5 per 1,000 successful requests, or $1 with Fast Search, with page text extracted into the snippet and no token bill. One POST /search can carry up to five queries and still be one billing unit. The rate-limit page counts those five as five query units against a 50-unit-per-second cap. An empty successful response is billed. A 429 is not. The Agent API’s web_search tool is $0.0025 a call, which is $2.50 per 1,000, or $0.001 on Fast Search. perplexity/sonar on the Agent API models page is $0.25 per million input tokens and $2.50 per million output tokens. The tool fee sits on top of that.

OpenAI’s pricing page says web search, including images, is $10 per 1,000 calls plus search-content tokens at the model rate. The note under the table says those tokens are retrieved from the index and fed to the model with the prompt. For gpt-4o-mini and gpt-4.1-mini, the non-preview tool bills that content as a fixed block of 8,000 input tokens a call. At the $0.15 per million input rate on the same page, 8,000 tokens are $0.0012. Add the $0.01 tool fee and the search, before any output tokens, is $0.0112 on those two models. That multiplication is this page’s. Do not copy it onto a model whose search content is not a fixed 8,000. The guide’s rate limit is the model’s tier, not a separate search quota.

Claude’s tool page and the pricing page use the same sentence: $10 per 1,000 searches, plus standard token costs for search-generated content. Each search counts as one use no matter how many results come back. An error is not billed. Results retrieved in the turn, and in later turns, count as input tokens. The docs point the rate limit at the Claude Console. They do not print one public number, so this page does not invent one.

Firecrawl’s search docs price a search at 2 credits per 10 results, rounded up. The API reference says that without scrape options the response is the URL, the title, and the description. Markdown is scrapeOptions. The pricing page’s Hobby plan is $19 a month for 5,000 credits. Dividing that by 5,000 and multiplying by 2 credits is $0.0076 a search, or $7.60 per 1,000. Standard is $99 for 100,000 credits, which is $1.98 per 1,000 on the same steps. The free plan is 1,000 credits, so 500 searches of up to 10 results. Firecrawl’s own search page states those two dollar figures. The division here uses the pricing-page inputs so the quotient can be checked.

SerpAPI’s FAQ says only successful searches count, and that 100 results cost the same one credit as an empty set. Cached, errored, and failed searches do not. The official MCP server describes more than one engine. The pricing page’s unit is the successful search, not a per-URL scrape.

A four-search task, tool fee only

An agent that searches four times, and then reads the results on your own model bill, pays the tool fee four times. Tokens are not in this column. Using the entry paid price, and assuming a SerpAPI or Firecrawl plan is fully used:

At 100,000 such tasks a month, that is 400,000 searches. Serper Starter’s 50,000 credits do not cover it. Scale, at $0.50 per 1,000, is $200, if the credits are used inside six months. SerpAPI Starter cannot cover it either. Production’s $10 per 1,000 would be $4,000. Cloud 1M’s $3.75 per 1,000 cannot cover 400,000 inside a 1,000,000 plan without a second question about the rest of the month, and this page does not price a partial Cloud plan. The native $10 tool fee alone would be $4,000, and the context-window tokens would be the larger line on any model that actually reads the pages.

That is the cost-per-task check. A $10 tool that writes the answer can still be cheaper than a $1 tool plus a long page dumped into a frontier model. It is not cheaper when the harness only needed ten URLs.

Which ones are an MCP server?

An MCP server lets a harness call the tool without you writing the HTTP client. It does not change the price. The key still bills the account.

Serper has third-party MCP servers. They are not on Serper’s pricing page, and this page does not recommend one. The Exa MCP repository says the hosted server works anonymously inside a rate limit, and that an API key or OAuth is how you get a higher limit and Exa Agent. Tavily’s free plan is 100 requests a minute, shared by every key in the account, and a search counts as one request whatever the depth.

What this means when you wire the harness

Put the native tool on the path where the product is a cited answer and you do not want a second vendor in the loop. Budget $10 per 1,000 searches, then budget the tokens. On OpenAI, confirm the request says web_search. A leftover web_search_preview on a non-reasoning model is the $25 row. On gpt-4o-mini and gpt-4.1-mini, add the fixed 8,000 input tokens, which are $0.0012 at the published mini rate, on top of the $0.01 call.

Put Serper on the path where the agent needs Google’s result page and your own code will decide what to open. The entry paid price is $1 per 1,000, the free start is 2,500 queries, and the credits expire in six months, so a $3,750 Ultimate pack is only $0.30 per 1,000 if you actually spend 12.5 million credits in time.

Put Exa Lite on the path where ten cached results are enough and you do not want a separate extract call. $4 per 1,000, 10 queries a second, $10 a month free.

Put Tavily on the path where you want snippets, a no-card month of 1,000 basic searches, and an MCP URL. Do not pay for advanced, at 2 credits, unless you need the longer chunks.

Do not put SerpAPI Starter on an agent because the logo says search. Twenty-five dollars buys 1,000 searches, and the hour stops at 200. Production, at $10 per 1,000 if you use all 15,000, is the first SerpAPI row that matches the native tool fee, and it is the row where the Legal Shield starts. Buy it for the engines and the shield, not because it was the first SERP API in the docs.

Brave is the high-throughput middle: $5 per 1,000, 50 a second, and the monthly $5 credit needs a card. Firecrawl is the wrong row if you only wanted snippets, and the right row if the agent’s next tool needs the page in markdown and you are willing to pay the scrape credits on top of the 2-credit search.

The app products that already search for a person sitting in a chat are still Best AI agent for web search. They are not this rate card.