What the built-in web-search tool adds to your bill when the model browses — priced per search across OpenAI, Anthropic and Gemini.
Turning on web search adds a per-search fee and a per-search token load. For a grounded chatbot doing millions of queries, both add up fast. Cap searches per request and cache stable context. Estimate the base generation with the LLM token cost calculator and full apps on the AI agent cost calculator.
The LLM Web Search Tool Cost Calculator estimates the extra monthly bill that a model's web-search tool adds on top of your normal token costs. You enter your requests per month, the average number of searches per request, and the provider's price per 1,000 searches, and it multiplies these to find your search-fee total. It then accounts for the extra result tokens per search — the retrieved snippets fed back into the model — priced at your input rate per 1M tokens. The two main cost drivers are search volume (requests times searches) and how many result tokens each search injects.
This matters because search fees and injected tokens are billed separately from your base prompt, so they quietly inflate spend. If costs run high, the most direct lever is reducing searches per request — cache or reuse results, and trim how many results you pull, since each one adds both a search fee and input tokens on every call.
Providers bill the built-in search tool per search — commonly around $10 per 1,000 searches (roughly $0.01 each) — and that is on top of tokens. A request that fires two searches pays two tool fees plus the tokens for whatever the searches return, which is often the larger cost.
Each search injects fetched snippets or pages back into the context, and you pay input-token price to process them every time. At a few thousand tokens per search across many requests, that token cost frequently exceeds the per-search tool fee itself. Limiting searches per request and trimming returned content is the main lever.
Cost at different volumes:
| Volume | Monthly | Yearly |
|---|