Cloudflare Workers AI and AI Gateway Are Now One Control Plane

Cloudflare shipped Workers AI and AI Gateway into a single control plane today. One API binding. One observability layer. One billing system. If you are using env.AI.run() in a Worker right now, you can add full gateway telemetry with a one-line change. If you want to route a request to OpenAI or Anthropic through the same binding you use for LLaMA, that works too. The wall between Cloudflare-native inference and external providers is down.

The Before State (and Why It Was Annoying)

Workers AI and AI Gateway have always done different things. Workers AI ran models on Cloudflare’s GPU infrastructure and exposed them via env.AI.run(). AI Gateway proxied requests to external providers and added an observability layer — but required separate setup, a different code path, and a different billing account. Developers building multi-provider AI systems on Workers ended up maintaining two parallel systems: one for their Cloudflare-hosted models and one for everything else. That split made observability fragmented and cost tracking tedious. Today’s update collapses it.

The One-Line Migration

This is the piece most relevant to anyone already using Workers AI. Adding AI Gateway observability to your existing inference calls requires exactly one additional parameter:

// Before: no observability
const response = await env.AI.run(
  '@cf/meta/llama-3.3-70b-instruct',
  { messages: [{ role: 'user', content: 'Hello!' }] }
);

// After: full observability, one line added
const response = await env.AI.run(
  '@cf/meta/llama-3.3-70b-instruct',
  { messages: [{ role: 'user', content: 'Hello!' }] },
  { gateway: { id: 'default' } }
);

Cloudflare auto-creates a default gateway on first authenticated request — no dashboard configuration required. Once enabled, every request logs request and response payloads, token counts per model, cost attribution, latency breakdowns, and error rates. Existing env.AI.run() calls without the gateway parameter continue to work unchanged. This is additive, not breaking.

External Providers Through the Same Binding

The more significant change is that AI.run() now works for external providers. Previously, calling OpenAI from a Worker meant a separate fetch() call, its own authentication setup, and manual AI Gateway proxy configuration. Now you specify the model string and the gateway ID:

// Call OpenAI’s gpt-5-mini from the same Workers AI binding
const response = await env.AI.run(
  'openai/gpt-5-mini',
  { messages: [{ role: 'user', content: 'Hello!' }] },
  { gateway: { id: 'my-gateway' } }
);

Switching providers is now a change to the model string. The catalog covers 70+ models across 12 providers, mixing Cloudflare-hosted open-source models (Meta, Mistral, Google Gemma, Microsoft Phi) with external frontier models from OpenAI, Anthropic, Kimi, and others. Multimodal capabilities — image, video, speech — are included.

Unified Billing and the Rate Limit Incentive

AI Gateway credits now apply to Workers AI inference. One credit balance covers both. Cloudflare is also offering a concrete incentive to make the switch: developers who enable unified billing get elevated rate limits on Workers AI models. If you were running into Workers AI rate caps before, this is the straightforward fix — set up unified billing and those limits increase without any other code changes.

Model-First Routing: Specify the Model, Not the Provider

Beyond the billing consolidation and code simplification, the architectural change that matters most for agent-heavy workloads is model-first routing. Instead of hardcoding a provider URL, you specify the model you want. Cloudflare handles provider selection and automatic failover: if the primary provider is at capacity, the AI Gateway routes to a backup serving the same model, transparently, while respecting Zero Data Retention requirements. On the roadmap is Smart Routing — an internal classifier that analyzes the prompt itself and routes simple classification or extraction tasks to cheaper Workers AI models while sending complex reasoning tasks to frontier models automatically. That feature is not live yet; factor it into your architecture plans, but do not design production systems around it until it ships.

The Agents Week Thread

This is Day 5 of Agents Week. Cloudflare shipped Kitesurf (a browser for AI agents), Cloudflare OS (an agent workspace), and Wallets (agent payments) this week. Every one of those products runs inference. The unified AI control plane is what ties them together — the shared inference backbone for the agentic computing stack Cloudflare is assembling. The full announcement is on the Cloudflare blog. Understanding the control plane helps make sense of why the other pieces fit together the way they do.

What to Do Now

• Add { gateway: { id: 'default' } } to your existing AI.run() calls — instant observability, no setup, non-breaking change
• Set up unified billing via the AI Gateway dashboard — converts existing credits to cover Workers AI inference and unlocks elevated rate limits
• Browse the model catalog — 70+ models now accessible from a single binding; if you have been using separate fetch calls for OpenAI or Anthropic, consolidate to AI.run()