Cloudflare AI Gateway review

Cloudflare AI Gateway is a hosted gateway and control plane that sits between an AI application and model-provider APIs. Instead of each application calling providers directly, requests can pass through a Cloudflare gateway endpoint, where teams can configure routing, caching, rate limiting and fallback behavior. The service also collects request logs and usage analytics.

That position in the stack makes it an operational layer, not a model, SDK or self-hosted proxy. Cloudflare describes routing by latency, cost or availability, and says rules can be adjusted from the dashboard or API without redeploying. Its logs can include prompts, responses, provider, status, token usage, cost and duration. The practical use is centralizing controls and visibility across provider calls. Teams still need to decide which providers and models to use, and should check how gateway behavior interacts with each provider's API and billing.

The integration surface is an HTTP gateway endpoint: Cloudflare shows a curl request to https://gateway.ai.cloudflare.com/v1/{account_id}/{gateway_id} and JavaScript examples for Workers AI and gateway configuration. A Wrangler JSONC example adds an AI binding to a Worker. For OpenAI-compatible calls, the example imports the openai package and sets its baseURL using env.AI.gateway(...).getUrl("openai").

Cloudflare says AI Gateway is compatible with the OpenAI SDK and AI SDK. The examples also show caching headers, fallback requests and methods such as patchLog and getLog for working with log records. The product page links to documentation and shows dashboard configuration. It does not name a CLI or a broader list of supported languages. Confirm SDK support, provider compatibility and endpoint-specific behavior in Cloudflare's docs before choosing an integration path.

Cloudflare's product page invites teams to start building for free, but the pricing page does not list AI Gateway-specific prices or billing terms. There is no published starting price, paid plan price, request allowance, token allowance, seat charge or overage rate to use for a production estimate. The free entry point is an advertised starting option, not a published quota with defined limits.

The homepage promotes unified billing and says teams can manage costs across providers through one gateway. It does not explain which charges are consolidated, how provider usage is billed, or whether gateway features add charges as request volume grows. Caching can reduce repeated provider calls, but savings depend on the application and cache behavior; it does not establish a gateway price. Before rollout, ask Cloudflare for the applicable billing model, free-tier limits, rate limits, cache-related charges, and any provider billing terms. Until those details are confirmed, compare the total expected cost with direct provider billing and other gateway options rather than assuming the free setup covers production use.

Cloudflare presents AI Gateway as a service running on its infrastructure, with requests sent to a Cloudflare gateway endpoint. Its product page does not publish a source-code license or describe a self-hosted edition, so teams should treat the documented deployment model as Cloudflare-hosted unless Cloudflare confirms otherwise.

The page describes logs and API methods for reading or updating individual log records, but it does not document bulk data export or a migration path. Moving away would mean changing the application’s gateway endpoint and replacing any Cloudflare-specific configuration, such as Workers AI bindings or logging calls. Provider-facing OpenAI-compatible requests may help keep some application code portable, but teams should test provider compatibility and preserve their own logs and routing configuration. Ask about export formats and deletion terms before routing sensitive traffic.

Cloudflare says the gateway runs on its global infrastructure and provides caching, fallback routing, rate limiting and security guardrails. Those are useful controls, but the product page does not publish gateway-specific request limits, availability targets, log retention periods or detailed compliance terms. Confirm thresholds, fallback behavior, streaming support, cache controls and incident handling against the requirements of your application.

Logs can contain prompts and model responses, so check what is retained, who can access it and whether sensitive fields can be filtered before sending production traffic. Security controls are advertised, but teams should verify how guardrails work for their providers and data flows. Reddit users describe routing Ollama calls through the gateway to inspect token usage and response time; another thread discusses using it with several OpenAI-compatible models. A separate comparison describes Cloudflare as quick to set up for teams already in its ecosystem. These are individual reports, not service guarantees.