Quick answer
Cloudflare now provides a unified access path across Workers AI and supported third-party providers: one AI binding and shared /ai/ REST endpoints, with AI Gateway applying observability, logging, caching, security, rate, retry, and billing controls.
This is a control-plane convergence, not proof that every feature or provider behaves identically. Workers AI still hosts models on Cloudflare infrastructure; AI Gateway manages inference traffic and can route to supported providers.
What is current
| Surface | Current source-backed behavior |
|---|---|
| AI binding | env.AI.run() supports Workers AI and supported third-party models |
| REST | Shared Cloudflare-authenticated /ai/ endpoints cover models and common SDK shapes |
| Default gateway | Created automatically on the first authenticated request when default is specified |
| Observability | Request volume, errors, latency, tokens, cost, and optional payload logging |
| Controls | Gateway caching, rate limiting, retries, security, and billing configuration |
| Billing | Prepaid AI Gateway credits can pay for Workers AI when Unified billing is enabled |
Use named gateways when applications, environments, teams, data rules, budgets, or retention need separate control boundaries. A default gateway is a convenient starting point; it is not automatically the right production segmentation.
What remains roadmap
Cloudflare describes model-first routing as coming soon, with a hoped-for pilot in coming months. It would select among providers hosting the requested model and support failover. Smart routing, where the gateway classifies a task and selects a model, is described as an internal pilot.
Do not promise automatic provider failover, model equivalence, Zero Data Retention routing, or zero-configuration task routing until current documentation establishes availability for the account and request path. Read the routing roadmap guide for planning boundaries.
Production control review
Before migration, inventory each binding and endpoint, model/provider, authentication method, gateway, payload class, log setting, cache key, retry policy, rate and spend rule, billing mode, data region, and fallback behavior. Validate whether prompts and responses may enter logs, who can query them, and how retention and deletion work.
Compare old and new paths with shadow or non-sensitive tests. Check output shape, streaming, errors, token accounting, latency, retries, cache behavior, duplicate risk, pricing, and limit enforcement. Keep rollback until monitoring confirms equivalence.
Use the migration guide for endpoint work, default gateway guide for logging, and readiness checker for a local gate.
Frequently asked questions
Are Workers AI and AI Gateway now the same product?
No. Cloudflare has unified model-access entry points and control-plane capabilities while Workers AI remains hosted inference and AI Gateway manages traffic, observability, controls, and billing.
Can one binding call Workers AI and third-party models?
Yes. Cloudflare documents the AI binding through env.AI.run() for Workers AI and supported third-party models, with gateway configuration selecting the control path.
Are model-first and smart routing available now?
The announcement labels model-first routing as coming soon and hoped for pilot in coming months, while smart routing is described as an internal pilot.
Official sources
- Cloudflare: Workers AI and AI Gateway unification
- Cloudflare changelog: Unified model access and billing
Source check: August 12, 2026. Recheck supported providers, endpoints, models, controls, billing, limits, logging, and roadmap status before use.