Unifying Workers AI and AI Gateway into a single AI control plane

Cloudflare announced the unification of its AI Gateway and Workers AI into a single control plane, allowing users to connect to any model provider—including Workers AI—while managing observability, billing, security, and logging from one interface. Previously separate products, AI Gateway and Workers AI now share the same entry points via Workers bindings and a unified REST API endpoint (/ai/), eliminating the need to choose between them. Users can automatically inherit AI Gateway features like logging and observability without prior setup by using a default gateway ID, streamlining initial deployment. The change reduces complexity for teams managing inference traffic, as all requests are logged with full payloads, token counts, and cost attribution from the first authenticated interaction.

The update introduces unified billing, enabling users to apply AI Gateway credits across external providers such as OpenAI and Anthropic, as well as Workers AI, with elevated rate limits for Workers AI when using the unified billing path. The Cloudflare AI Gateway dashboard now provides detailed visibility into every request, including latency, token usage, error rates, and full prompts and responses, addressing debugging and auditing needs. This unified approach allows teams to scale without managing separate gateways, with the option to create custom named gateways for advanced use cases like traffic splitting or custom caching rules.

Cloudflare is shifting from provider-first routing to model-first routing, where users specify the model they need (e.g., Kimi K2.7 Code) rather than the provider, allowing the control plane to handle provider selection, failover, and load balancing automatically. This approach improves resiliency by transparently routing traffic to alternative providers if the primary one experiences issues, without requiring application-level retries. The system prioritizes model quality and compliance with requirements such as Zero Data Retention (ZDR), ensuring consistent output standards across vetted providers.

The next phase of routing will introduce intelligent model selection, where a classifier running on Workers AI analyzes prompts to predict task type, complexity, and context, then maps them to the optimal model from a curated pool without manual configuration. While users can still specify exact models, the zero-config path aims to deliver better economics and performance by automating routing decisions. Cloudflare is piloting this internally and plans to expand testing before release, with early access available to existing Workers AI users via the default gateway setup.