Quick answer
Cloudflare’s unified Workers AI and AI Gateway entry points, default-gateway observability, and unified billing are current. Model-first routing and smart routing are roadmap surfaces in the August 7 announcement.
Model-first routing is labeled “coming soon,” with a hoped-for pilot in coming months. Smart routing—classifying a request and choosing a model—is described as an internal pilot. Do not build a production availability promise from the example request or future architecture.
Current versus roadmap
| Capability | Status in the announcement |
|---|---|
| Unified AI binding and REST access | Current |
| Default gateway observability | Current |
| AI Gateway credits for Workers AI | Current |
| Provider selection by application | Current responsibility |
| Same-model provider failover | Coming-soon model-first roadmap |
| Automatic task-to-model selection | Internal smart-routing pilot |
| Guaranteed model/provider equivalence | Not established |
The announced model-first idea starts with a requested model, then selects among vetted providers hosting it based on availability and constraints. Smart routing goes further by classifying task type and complexity, then choosing from a curated model pool.
Evaluate before depending on routing
Define acceptable providers, regions, data handling, Zero Data Retention requirements, model versions, output variance, latency, cost, capacity, and failure modes. Same weights do not automatically establish identical quantization, serving stack, moderation, tool behavior, context limits, or performance.
For any future pilot, log the selected model and provider, reason or policy, routing decision, fallback, latency, tokens, cost, and result. Test provider outage, rate limiting, regional constraint, model drift, incompatible response shape, and recovery to the preferred path.
Keep an application-level way to pin a provider or model when regulated data, deterministic evaluation, contractual terms, or quality thresholds require it. Define whether failover is allowed before the request, not after an outage.
Avoid placeholder implementation
Do not reproduce future example endpoints as guaranteed production APIs. Wait for current developer documentation to define enrollment, endpoint, supported models and providers, constraints, observability, pricing, and service-level behavior.
The safest work now is to centralize model policy, create portable evaluations, label provider constraints, and make routing decisions observable. Read the current control-plane guide and use the readiness checker for present-day migration.
Frequently asked questions
Is Cloudflare model-first routing available now?
The August 7 announcement labels it coming soon and says Cloudflare hopes to pilot it in the coming months; verify current documentation before relying on it.
What would model-first routing do?
Cloudflare describes selecting among providers that host the requested model, including load balancing or failover when a provider has capacity or availability problems.
Is smart model selection available now?
Cloudflare describes prompt classification and automatic model selection as an internal pilot being tested and iterated before release.
Official source
Source check: August 12, 2026. Recheck pilot availability, endpoints, providers, models, constraints, telemetry, pricing, and production support.