Cloudflare has added a Web Search API beta to AI Gateway, giving agents and AI applications a way to search the live internet through the same gateway layer already used for model traffic. The feature is new, narrow in surface area, and still in beta, but it points at a larger shift: search is becoming a billable AI tool, not just an application-side integration.
The company announced the beta on October 2, 2026. At launch, Cloudflare lists Ceramic.ai, Exa and Linkup as selectable search providers. Developers can send search requests through AI Gateway, have those requests appear in gateway logs, and pay for them through AI Gateway credits at each provider’s listed API price. Cloudflare says it is not adding an additional markup. Teams can also use bring-your-own-key provider credentials where supported.
That combination matters because it joins three concerns that are often handled separately: tool routing, observability and billing. For developers building retrieval-augmented agents, research assistants, support bots or automated sales workflows, the question is no longer only which model generated the answer. It is also which search provider was called, under whose credentials, at what price, and whether the resulting spend can be tied back to a customer, team, project or API key.
What changed
Until recently, many AI gateway discussions centered on model calls: OpenAI-compatible endpoints, provider failover, token accounting, latency, caching and request logs. Cloudflare’s Web Search API beta adds a different kind of transaction to that control plane. A search call may support a model response, but it is not itself a model inference request. It has different pricing, different provider semantics and different governance questions.
Cloudflare’s launch makes the search provider a configurable gateway choice rather than a hard-coded application dependency. Developers can choose among the listed providers, route the call through AI Gateway, and see it alongside other AI traffic. If the customer uses Cloudflare-managed billing, the charge runs against AI Gateway credits. If the customer brings a provider key, the gateway can still sit in the path for routing and logging.
That is a meaningful product boundary. Search is no longer an invisible helper hidden behind an agent framework. It becomes a first-class event in gateway records. For a multi-model API gateway for teams, that is the difference between “the agent was expensive” and “this workflow spent money on model tokens, cached context and six live search calls.”
Why gateway teams should care
The immediate audience is developers already using Cloudflare AI Gateway, especially those building applications that need fresh information. Search can help ground responses in live data, but it also introduces another unit of cost and another source of provider dependency. If a support agent searches documentation, a compliance assistant checks current policy pages, or a market-monitoring workflow scans the web repeatedly, search charges can become operationally important even when token usage looks modest.
This is the same pattern now appearing across agent infrastructure. Hosted shell tools, browser use, code review agents and retrieval services are all turning into metered operations around the model. Tokens still matter, but they are no longer the whole ledger. A usable gateway has to explain spend across the complete workflow, not just the final completion.
For businesses, that affects chargeback and margin. A reseller or internal platform team may need to decide whether search spend is passed through at provider list price, marked up, bundled into a plan or blocked by default. A finance team may want to know whether a customer’s invoice increased because of larger prompts, more output tokens, cache behavior or web search volume. Product teams may want usage reports that distinguish background enrichment from user-visible generation.
That creates a practical opening for stronger unified AI API billing. The billing model has to cover tokens, cached tokens and tool calls without flattening them into a single vague “AI usage” line. Otherwise, customers cannot tell which part of an agent system is driving cost.
The developer consequences
Developers will need to treat web search as part of the application’s API design, not just as a prompt enhancement. The search provider should be visible in configuration. Credentials should be scoped. Logs should preserve enough information to debug provider behavior and customer spend without retaining unnecessary sensitive data. Error handling should distinguish a model failure from a search provider failure, because the user experience and retry strategy may be different.
There is also a routing question. Ceramic.ai, Exa and Linkup are not interchangeable in every use case. Search quality, coverage, latency, result format and pricing can all affect downstream model behavior. A gateway that can route between providers gives developers flexibility, but it also requires policy: which provider is approved for which tenant, which regions or data classes are allowed, and whether BYOK credentials are required for some customers.
For Model Gate and similar products, the practical connection is direct. If customers are building agent services on top of a gateway, they will eventually ask for search usage to appear next to model usage in the same analytics surface. Partner APIs may also need to expose search-call records so downstream platforms can bill their own customers accurately. Treating search as a first-class billable tool would fit naturally beside API-key management, team controls and usage analytics.
That does not mean every gateway needs to copy Cloudflare’s exact provider list. The more important point is structural: once live search is available through a gateway, tool metering becomes part of the gateway contract. A team AI API spend control that ignores tool calls will understate real agent cost.
What remains uncertain
The beta status leaves several open questions. Cloudflare has named the initial search providers and described the billing approach, but production teams will still need to evaluate provider-specific performance, result quality, data handling commitments and operational limits. Public documentation confirms the beta and the basic pricing path, but it does not settle whether one provider is better for all workloads or how the feature will evolve before general availability.
It is also unclear how quickly developers will standardize around gateway-mediated search rather than direct provider integrations. Teams with existing retrieval stacks may keep search inside their own infrastructure. Others may prefer gateway routing because it reduces integration work and centralizes logs. The decision will likely depend on governance needs as much as search relevance.
Still, the direction is clear enough. AI gateways are absorbing more of the agent runtime around the model. Cloudflare’s Web Search API beta is not just a search feature; it is another sign that the control plane for AI applications now has to meter, route and explain tools as carefully as it handles models.