On October 2, 2026, Cloudflare added a Web Search API to its AI Gateway, in open beta, with three search providers priced from $0.25 to $7 per 1,000 requests. The search tools built into the big model APIs cost $10 to $14 per 1,000 searches on current models before token charges, and $35 per 1,000 prompts on Gemini 2.5. This page puts Cloudflare’s launch beside four model-native tools and seven standalone search APIs, with each price taken from the vendor’s own page.
- 01A 56-fold spreadCheapest-tier prices run from $0.25 to $14 per 1,000. The lowest is Ceramic.ai, Cloudflare’s default provider.
- 02Model tools cost moreOpenAI and Anthropic charge $10 per 1,000 searches and then bill the results as model tokens.
- 03Units differPer search, per query, per prompt and per credit are four different things. Compare on your own traffic.
- 04Terms decide reuseGoogle and Microsoft attach display and use rules to grounded results. Raw-result APIs mostly do not.
01 — The releaseWhat Cloudflare launched
Cloudflare’s announcement adds web search as one more thing AI Gateway can call, alongside the model providers it already routes to. A request goes to one endpoint or a Workers binding and names a provider. At launch there are three: Ceramic.ai, which is the default, Exa and Linkup. Every provider returns the same result shape, a title, URL and description per result, so switching providers is one parameter.
Cloudflare says it bills at each provider’s list price with no markup, and the providers page gives $0.25 per 1,000 for Ceramic.ai, $5 for Linkup and $7 for Exa. The Exa price matches Exa’s direct list price, and the Linkup price sits at the low end of Linkup’s direct $5 to $6. Costs come out of AI Gateway credits, or a team can bring its own provider key and be billed by the provider. Two limits matter: at most 10 results per request, and a native server tool that a model can call on its own is marked “coming soon”. For now, a developer defines a search function for the model and calls the API from it.
02 — ContextTwo kinds of search tool
The products on this page do two different jobs, and the price gap mostly follows the job.
Model-native search
The model decides when to search, reads the results and writes a cited answer. You get the answer, not a reusable list, and only with that vendor’s models.
Raw-result search API
Your code sends a query and gets ranked results with snippets. Any model can read them, and you decide what to keep.
Model-native search is less code: one flag on a request. A raw-result API is more work but portable, and it lets an agent search once and feed the same results to a cheaper model. Our comparison of managed retrieval services covers the other half of the problem: searching your own documents rather than the web.
03 — The dataPrice per 1,000 searches
Each row gives the list price for the default or cheapest agent-search tier, what else is billed on top, and any free allowance. Volume discounts and enterprise contracts are left out.
| Service | Per 1,000 | Free allowance | Also billed |
|---|---|---|---|
| Cloudflare, Ceramic.ai (default) | $0.25 | None stated | Nothing from Cloudflare; your model bills its own tokens |
| Cloudflare, Linkup | $5.00 | None stated | Nothing from Cloudflare |
| Cloudflare, Exa | $7.00 | None stated | Nothing from Cloudflare |
| OpenAI web_search tool | $10 | None | Search content tokens at the model’s rates |
| Anthropic web search tool | $10 | None | Search results billed as input tokens |
| Google grounding, Gemini 3.x | $14 per 1,000 queries | 5,000 queries a month on the paid tier | Model tokens; one prompt can run several queries |
| Google grounding, Gemini 2.5 | $35 per 1,000 prompts | 1,500 prompts a day on the paid tier | Model tokens |
| Microsoft Grounding with Bing | $14 | None | Model tokens in Foundry |
| Brave Search API | $5 | $5 of credits a month | Nothing |
| Exa, direct | $7 (Fast or Auto) | $10 of credits a month | $1 per 1,000 for each result above 10 |
| Tavily | $8 basic, $16 advanced | 1,000 credits a month | Nothing |
| Perplexity Search API | $5 ($1 Fast) | None found | No token costs |
| Parallel Search API | $1 to $5 by mode | Stated two ways on its page | $1 per 1,000 extra results |
| Linkup, direct | $5 to $6 | 4,000 queries | Nothing |
| Ceramic.ai, direct | $0.25 | 1,000 credits | Nothing |
Headline list price per 1,000 searches, US dollars, lower is cheaper
Vendor pricing pages, read October 3, 2026. Cheapest agent-search tier per vendor. OpenAI, Anthropic, Google and Microsoft also bill model tokens on top. Google’s figure is the Gemini 3.x price per search query after the free allowance.04 — The catchWhy the prices do not compare directly
A price per 1,000 hides what one unit buys. Four differences change the real bill.
- Tokens on top. OpenAI’s web search tool and Anthropic’s web search tool bill the retrieved content as model input, so the real cost per search depends on the model’s input price. For two small models, OpenAI bills a fixed block of 8,000 input tokens per call, and it caps the search context at 128K tokens.
- Query or prompt. Google’s Gemini API pricing charges Gemini 3.x per search query the model runs, and one prompt can run several. Gemini 2.5 charges per grounded prompt, however many searches it takes.
- Results per unit. Cloudflare returns at most 10 results a request; Exa charges $1 per 1,000 for each result above 10. A team that needs 30 results pays very differently on each.
- Credits. Tavily bills credits: one for a basic search, two for advanced. Anthropic does not bill a search that fails, and counts one use per search however many results it returns.
05 — The dataLimits and data terms
Price is only half the choice for an agent in production. Rate limits decide whether it survives a busy hour, and data terms decide whether a regulated team can use it at all.
| Service | Rate limit | Results | Data and use terms |
|---|---|---|---|
| Cloudflare Web Search API | Not published | Up to 10 | Cloudflare lists zero data retention for Ceramic.ai and Linkup, not Exa |
| OpenAI web_search | The model’s tier limits | Not exposed | Search context capped at 128K tokens; live search not covered by a BAA |
| Anthropic web search | Not published | Not exposed | Basic version eligible for zero retention; not on Amazon Bedrock |
| Google grounding | Not published | Not exposed | Prompts stored 30 days; Search Suggestions must be shown; no caching or training on results |
| Grounding with Bing | 150 a second; 1M a day | Not stated | Azure AI Foundry or Azure AI Search only; Microsoft’s DPA does not apply |
| Brave | 50 a second | 20 | Storing results needs a plan with storage rights |
| Exa | 10 a second, up to 25 | 100 | Zero retention on Enterprise |
| Tavily | 100 a minute (dev), 1,000 (production) | 20 | No retention option found |
| Perplexity | 50 query units a second | 20 | Zero retention stated for its chat API only |
| Parallel | 600 a minute | 20 | No retention option found |
| Ceramic.ai | 20 a second; 50 on Pro | 20 | English web pages only; retention terms on request |
Cloudflare’s provider table marks Ceramic.ai and Linkup as zero data retention. Bought directly, Linkup offers that only on its Enterprise plan, and Ceramic.ai asks customers to contact it. The same provider can carry different terms depending on who sells it to you. Check the terms on the route you will actually use.
One older route has gone. Microsoft retired its public Bing Search APIs on August 11, 2025; its replacement, Grounding with Bing, works only inside Azure AI Foundry and Azure AI Search, and its outputs cannot be used directly in other applications. Agents that read community sites face a similar squeeze, as our note on Reddit’s API and RSS deadlines sets out.
06 — Practical implicationsWhich one to use
Whatever the shortlist, run the same 200 real queries from your agent’s logs through two or three options and compare the bill and the answers. Vendors publish index sizes and latency figures of their own, and we have not reproduced any of them. For teams that want this tested and wired into an agent, our AI transformation work covers tool selection, cost controls and evaluation. The Perplexity row has more context in our guide to its Agent API.
07 — MethodMethod and as-of date
A comparison of published prices and documented limits. Nothing on this page was load-tested by Digital Applied.
- What was collected
- List price per 1,000 requests for the default or cheapest agent-search tier, charges billed on top, free allowance, rate limit, results per query and data or use terms, for 15 routes across 12 vendors.
- Sources
- Each vendor’s own pricing page, API reference or terms. Cloudflare’s announcement and Web Search documentation for the anchor. No third-party price lists or comparison sites were used.
- As-of date
- October 3, 2026. Pages with no visible date were checked against archived copies from before October 2, and no listed figure had changed. Perplexity’s pricing, Anthropic’s documentation, Ceramic.ai and some documentation subpages could not be checked that way.
- Units
- US dollars per 1,000 requests, except Google (per search query for Gemini 3.x, per grounded prompt for Gemini 2.5) and Tavily (credits converted at the pay-as-you-go rate). Rate limits as each vendor states them, per second or per minute.
- Exclusions
- Enterprise and volume pricing; Anthropic’s web fetch tool, which has no per-call fee; answer-generating plans such as Brave’s Answers tier; Exa’s Deep modes at $12 and $15 per 1,000.
- Limitations
- No latency, quality or index-size claim was tested; vendor figures for those are their own. Parallel’s page states its free allowance two ways. Cloudflare has not published rate limits for the beta.
- Refresh
- Re-read every pricing page monthly and when Cloudflare adds a provider or ends the beta. Correct figures in place with a dated note.
Price your own query log before you pick
Take a week of your agent’s real searches, count queries, results and tokens, and price that log on two or three rows from this page. The cheapest headline is rarely the cheapest bill, and the terms on reuse can rule out an option before price matters.