Quick answer
How much do built-in web search tools cost compared with a search API?
As of October 2026, OpenAI's and Claude's web search tools cost $10 per 1,000 calls, plus the retrieved content billed as input tokens. Gemini 3 grounding and Grounding with Bing cost $14 per 1,000. Standalone APIs cost $1 to $8 per 1,000 and return the results to you; LLMWebSearch costs $1.45/1K searches at 100,000 searches, with the top 3 pages included.
Built-in search tools and standalone search APIs side by side
| Tool | Fee per 1,000 | Retrieved content billed as tokens? | What you get back | Storage and display rules |
|---|---|---|---|---|
| OpenAI web_search (Responses API) | $10 per 1,000 tool calls | Yes, at the model's input rate; a fixed 8,000 tokens per call with gpt-4o-mini and gpt-4.1-mini | The model's answer with URL citations; optionally the list of consulted URLs | Citations must be visible and clickable |
| OpenAI web_search_preview, non-reasoning models | $25 per 1,000 tool calls | No: search content tokens are free | Answer with citations | Same as above |
| Anthropic Claude web_search | $10 per 1,000 searches | Yes, as input tokens, again on later turns | URL, title and page age; page content encrypted | Citations must be included when you show the output |
| Anthropic Claude web_fetch | No per-call fee | Yes, the fetched text as tokens | Full page text or PDF | Not checked |
| Google Gemini 3 grounding with Google Search | $14 per 1,000 search queries, after 5,000 free a month on the paid tier | Not stated | Citations and search suggestions, no snippets | Search suggestions must be shown; grounded text may be stored up to two years for limited purposes; no training |
| Google Gemini 2.5 grounding | $35 per 1,000 grounded prompts | Not stated | Grounding chunks (URI and title) | Same Google rules |
| Microsoft Grounding with Bing Search | $14 per 1,000 transactions | Model tokens billed separately | Results are not accessible to other applications | References shown in the exact form provided; no training |
| xAI web_search | $5 per 1,000 successful calls | Yes, all token types billed | Encountered URLs and inline citations | Not checked |
| Perplexity Agent API web_search | $2.50 per 1,000 ($1 fast); fetch_url $0.50 | Yes, search context billed as tokens | URL, title, snippet and dates | Not addressed |
| LLMWebSearch (standalone, ours) | $1.85/1K searches to $1.20/1K searches, top 3 pages included | No: you choose what to send to your model | Up to 20 ranked results and the top 3 pages as Markdown | No rule of ours against storing results; we do not train on queries |
| Tavily (standalone) | $7.50 to $5.00 per 1,000 basic searches | No | Results, optional raw page content and optional answer | Terms §6.5 allow training on customer input and output |
| Exa (standalone) | $7 per 1,000 auto searches, text of 10 results included | No | Results with page text | Terms grant Exa a licence to cache and store input and output |
| Brave Search API (standalone) | $5 per 1,000 requests | No | Results; LLM Context returns page chunks | Default terms forbid storing results and AI training |
Prices checked on
Model-provider prices from developers.openai.com, platform.claude.com, ai.google.dev, microsoft.com/bing/apis, docs.x.ai and docs.perplexity.ai; standalone prices from each vendor's pricing page. All read on 2026-10-22.
Why does a built-in tool cost more than its fee?
Because the pages it reads come back as input tokens. OpenAI bills search content at the model's rate, and with its mini models counts a fixed 8,000 input tokens per call. Claude bills search results as input tokens, and they count again on every later turn of the conversation.
The fee per call is only the visible part. The token part grows with the number of searches per task and the length of the conversation. Most built-in tools also decide for themselves how many searches to run.
What does a worked example look like?
This is an estimate with stated assumptions, not a quote. Take an agent that runs 100,000 searches a month.
- Built-in tool at $10 per 1,000: $1,000 in tool fees. Add the content tokens. At a fixed 8,000 tokens per call, that is 800 million input tokens. At an assumed $1.00 per million input tokens (a round figure, not a vendor price), that adds $800. Total: about $1,800.
- Standalone API: with LLMWebSearch, 100,000 searches cost $145, and each returns the top 3 pages. Your model still reads the text you pass it. If you send 1,500 tokens from each of 3 pages, that is 450 million tokens, or about $450 at the same assumed rate. Total: about $145 plus $450.
The gap comes from two places: the fee per call, and control over how much text reaches the model. With a standalone API you set read and max_chars, and you can drop pages your agent does not need.
When is a built-in tool the better choice?
When you want the least code. A built-in tool needs no second account, no API key and no tool definition: you switch it on in the model request. It also fits short chats with few searches, where the token add-on stays small.
Gemini's 5,000 free grounded queries a month on the paid tier cover many prototypes. Claude's web_fetch has no per-call fee, so fetching a known URL costs only tokens. And if you must show a provider's citations in its exact format anyway, the built-in tool already does it.
When is a standalone search API the better choice?
When you need the raw results, many searches per task, or the same tool across several models. A standalone API returns ranked results and page text to your code. You can cache them, rerank them, show them, or pass them to any model: OpenAI, Claude, Gemini or an open-weight model.
It also avoids being tied to one provider's rules. Grounding with Bing does not hand results to other applications. Google requires its search suggestions to be shown. With a standalone API, the results are yours to handle under that vendor's terms. Read those terms too: Brave's default terms forbid storing results, and Tavily's and Exa's terms reserve rights over your inputs and outputs.
Can I use a standalone search API with OpenAI or Claude models?
Yes, as a function tool. You describe a web_search function to the model, run the search when the model calls it, and return the results. Our OpenAI function-calling guide shows the full loop in Python, and the MCP server plugs the same search into Claude Desktop, Claude Code and Cursor. The cost guide helps size the monthly bill.
Questions about built-in search tools
How much does OpenAI's web search tool cost?
$10 per 1,000 calls for all models, plus search content tokens billed at the model's rate, on OpenAI's pricing page read on 2026-10-22. The older web_search_preview tool on non-reasoning models costs $25 per 1,000 calls, with content tokens free.
How much does Claude's web search tool cost?
$10 per 1,000 searches plus standard token costs for the search results, on Anthropic's pricing page read on 2026-10-22. The results count as input tokens again on later turns. Claude's web_fetch tool has no per-call fee.
What does Gemini grounding with Google Search cost?
For Gemini 3 and newer models, 5,000 search queries a month are free on the paid tier, then $14 per 1,000 queries. Gemini 2.5 models are billed $35 per 1,000 grounded prompts. Read on ai.google.dev on 2026-10-22.
Can I store results from Grounding with Bing?
Microsoft says the results are not directly accessible for use in other applications, and its terms require references to be shown in the exact form provided and bar training. A standalone API is the usual route when you need the results themselves.
Sources (10)
- OpenAI API pricing (built-in tools) — OpenAI (read 2026-10-22)
- Claude API pricing — Anthropic (read 2026-10-22)
- Claude web search tool — Anthropic (read 2026-10-22)
- Gemini API pricing — Google (read 2026-10-22)
- Grounding with Bing Search pricing — Microsoft (read 2026-10-22)
- xAI API pricing — xAI (read 2026-10-22)
- Perplexity API pricing docs.perplexity.ai/docs/getting-started/pricing — Perplexity (read 2026-10-22)
- Brave Search API terms of service api-dashboard.search.brave.com/terms-of-service — Brave Software (read 2026-10-22)
- Tavily terms of service tavily.com/terms — Tavily (read 2026-10-22)
- Exa pricing exa.ai/pricing — Exa (read 2026-10-22)