Last updated: October 2026
- Claude charges $10 per 1,000 searches and Gemini 3 models charge $14 per 1,000 search requests, both on top of tokens.
- Claude and OpenAI document domain filtering; the Gemini page does not: Claude takes allowed_domains or blocked_domains, OpenAI up to 100 of either.
- Claude caps cost with max_uses, and a search that errors is not billed.
- Gemini requires you to render Search Suggestions in your UI, which matters if grounded text is shown to readers.
SERP research is the slowest step in a content pipeline. A model that can search by itself removes the copy-paste between a rank tracker and a prompt.
The three providers built that differently. The differences show up in your bill, in how you restrict sources, and in what you must display.
How Do the Three Web Search Tools Differ?
| Feature | Claude | OpenAI | Gemini |
|---|---|---|---|
| Tool name | web_search | web_search (Responses API) | google_search |
| Price signal | $10 per 1,000 searches | Per tool call | $14 per 1,000 (Gemini 3+), 5,000 free a month |
| Search cap | max_uses | Not stated | Not stated |
| Domain control | Allow or block list | Up to 100 allow or block | Not listed |
| Citations | Always on, cited_text up to 150 characters | url_citation annotations | url_citation with start and end indices |
| Offline mode | Not applicable | external_web_access: false | Not listed |
Sources: Anthropic’s web search tool documentation, OpenAI’s web search guide and Google’s Grounding with Google Search page, read in October 2026. “Not stated” means the page does not say, not that the feature is absent.
Look up your OpenAI per-call rate on the pricing page before you plan a run. The guide says calls are charged, but the figure sits elsewhere.
How Do You Keep a Research Run from Blowing the Budget?
On Claude, set max_uses. Anthropic says simple factual queries typically use 1 to 3 searches, while comparative research can use 10 or more.
If Claude tries to go over the cap, the result block carries a max_uses_exceeded error. The API still returns HTTP 200, so check inside the response body.
Take 500 competitor briefs at 6 searches each. That is 3,000 searches, or $30 at Claude’s $10 per 1,000, before token costs.
The same volume on a Gemini 3 model is $42 at $14 per 1,000, minus the 5,000 monthly free queries on that plan. Multiple searches inside one API call count as separate billable requests.
Search results count as input tokens on every later turn of a conversation. A long research thread gets more expensive with each question, not just with each search.
Can You Restrict Which Sites the Model Reads?
For SEO research this is the control that matters. Without it, the model may cite a scraper site that copied the original.
Claude accepts allowed_domains or blocked_domains, never both, and returns a 400 if you send both. Entries are bare domains with an optional path, such as example.com/blog.
OpenAI allows up to 100 domains in either list, without the protocol, and includes subdomains automatically. That filter works with the Responses API web_search tool only.
A long Claude domain list can return a request_too_large error, per Anthropic’s documentation. Keep the list short and specific.
For fact-checking pipelines, allow only primary documentation domains. For competitor research, use a block list of aggregators instead.
What Must You Show Readers When You Publish Search-Grounded Text?
Each provider attaches conditions. Anthropic says citations must be included when API output is displayed directly to end users. OpenAI says inline citations must be clearly visible and clickable.
Google goes furthest. The Grounding page says the Search Suggestions HTML must be rendered according to its Terms of Service specifications.
If search output only feeds your own brief and a human rewrites it, these rules are about the draft stage. If grounded text goes live untouched, plan the UI work first.
Which Failure Modes Hit Search-Enabled Pipelines?
Claude can return stop_reason: "pause_turn" on a long-running search turn. You must send the paused assistant message back unchanged to continue.
Multi-turn conversations must also pass back each result’s encrypted_content exactly as received. If it is missing or modified, the request fails with a 400 validation error.
OpenAI caps search context at 128k tokens, whatever the model’s larger window. Plan summaries per source, not whole pages.
For context on cutting these token costs, see our guide to prompt caching for SEO pipelines.
When displaying API outputs directly to end users, citations must be included to the original source.
– Per Anthropic’s published web search tool documentation
Cap searches, restrict domains, and read the error field inside a 200 response. Claude gives the tightest cost and source controls; Gemini adds a display obligation.
Price a run as searches times the per-search rate, then add the input tokens the results create.
Frequently Asked Questions
How much does Claude web search cost?
Anthropic’s documentation lists $10 per 1,000 searches plus standard token costs. A search that errors is not billed.
How much does Gemini grounding cost?
Google lists $14 per 1,000 search requests for Gemini 3 and later, with 5,000 free queries a month. Gemini 2.5 and older models bill per prompt instead.
Can I use web search inside the Claude Batch API?
Yes. Anthropic’s documentation says batch web search is priced the same as regular requests and that Anthropic throttles search requests per organization, so large batches can run longer.
How many domains can I filter on OpenAI?
Up to 100 allowed domains or 100 blocked domains, per OpenAI’s web search guide. Subdomains are included automatically.
Does a search error return an HTTP error?
On Claude it does not. Errors such as too_many_requests or max_uses_exceeded arrive inside a 200 response, so parse the result block before trusting the text.
DesignCopy Editorial Team
English 
한국어 