Before consolidating Workers AI and AI Gateway, test the billing boundary

Cloudflare is bringing Workers AI and AI Gateway together as a single AI control plane. For teams that use more than one model provider, the attraction is straightforward: one routing surface with logging, caching, rate limits, and cost information instead of separate dashboards and credentials.

The operational decision is more specific than “move everything to the gateway.” A unified request path does not mean every model follows the same billing rule, retention setting, or enforcement behavior. Test those boundaries before making a gateway the default for production traffic.

What can be unified, and what cannot

Cloudflare documents an API that can route both third-party and Workers AI models through AI Gateway features such as logging, caching, and rate limiting. Its announcement says existing Workers AI users can route calls through a default gateway to obtain request logging, token tracking, and cost attribution.

Billing remains an important distinction. Cloudflare's Unified Billing documentation says that third-party provider models can use preloaded account credits, while Workers AI models with an `@cf/` prefix are billed through Workers AI pricing instead. Do not combine those figures into one assumed unit cost without checking the model and request path.

Run a small, representative gateway test

Choose one application and one non-sensitive workload that uses the same model classes as production. Send a known set of requests through a named gateway, then compare the gateway logs, token data, latency, provider bill, and Cloudflare bill. Include at least one third-party model and one `@cf/` model if your architecture uses both.

The goal is to verify accounting and controls, not to benchmark model quality. Keep the test scope small enough that a routing mistake, unexpected log setting, or budget block does not interrupt customer-facing work.

• Name the gateway for the application and environment; do not reuse an unexplained default for every workload.
• Record which requests use provider keys, Unified Billing, or Workers AI pricing.
• Compare token-based cost estimates with the provider's own invoice or usage dashboard.
• Check the logging and retention policy before sending prompts that may contain customer or employee data.
• Set a low spend limit and verify how the application handles a blocked request.
• Keep a direct-provider or previous routing path until the team accepts the test result.

Treat cost controls as protective, not exact accounting

AI Gateway spend limits can restrict requests when a budget is reached, using model, provider, or custom metadata dimensions. Cloudflare describes the calculation as a real-time estimate based on token usage and model pricing. That makes it a useful circuit breaker, but not a substitute for reconciling the provider invoice.

The same caution applies to dashboards. Cloudflare notes that cost metrics depend on returned token data and model names, and that provider pricing can change. Use gateway data to detect a trend or enforce a guardrail; use the contractual bill to close the financial record.

Decide on retention and routing deliberately

A shared control plane introduces a shared data-handling decision. Cloudflare supports Zero Data Retention as a gateway default or per-request override for Unified Billing. Confirm whether that setting, the selected logging configuration, and your application's prompt categories meet the policy your organization has already approved.

Cloudflare also describes future intelligent routing work as an internal pilot. Do not plan production behavior around an unshipped automatic router. Specify models explicitly where the choice affects quality, compliance, latency, or cost, and revise that decision only after an available feature has been tested against your workload.