Cloudflare Ai

Cloudflare Workers AI runs open-source model inference on Cloudflare's edge network — the same infrastructure that handles trillions of HTTP requests/month. No GPU provisioning: choose a model, make an API call, get inference results in milliseconds, pay per neuron (processing unit). Runs at 180+ PoPs worldwide, reducing inference latency for globally distributed applications. Model catalog: Llama 3.1/3.2/3.3, Mistral 7B, Whisper, SDXL, BGE embeddings, and 50+ others. For agents: use the REST API for any HTTP-accessible environment, or use the native Workers binding (ai.run()) for agents running as Cloudflare Workers — the fastest integration path. Free tier: 10k neurons/day. Confidence is docs-derived.

REST API: POST https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/{model_id} with JSON body ({messages} for chat, {text} for embeddings, {audio} for STT, {prompt} for text generation). Response: {result: {...}, success: true, errors: [], messages: []}. Chat: {messages: [{role: "system", content: "..."}, {role: "user", content: "..."}]}. Embeddings: {text: ["sentence 1", "sentence 2"]}. STT (Whisper): {audio: base64_encoded_audio}. Image gen (SDXL): {prompt: "...", num_steps: 20}. Streaming: add ?stream=true for SSE streaming of text generation. Workers binding: const ai = new Ai(env.AI); await ai.run("@cf/meta/llama-3.1-8b-instruct", {messages}).

API token auth: Authorization: Bearer {api_token} header. Token from Cloudflare Dashboard → My Profile → API Tokens (scope: Workers AI: Read). Account ID required in URL path (from Dashboard → account overview). HTTPS enforced. No OAuth2. Workers binding: automatically authenticated via Workers runtime (no explicit auth code needed in Worker). Rate limiting: per-account neuron budget per day; paid tiers for higher limits.

Standard HTTP status codes with JSON error bodies ({errors: [{code, message}]}). 401 for invalid token, 403 for missing AI permission, 400 for invalid model ID or malformed request body. Rate limiting: 429 when daily neuron budget exhausted. Streaming: SSE events with delta tokens; final event is [DONE]. Model availability: most models available globally; some large models have limited regional availability. Cold-start: negligible at Cloudflare edge (Workers run on V8 isolates, not containers). Cloudflare AI uptime tied to Cloudflare network (99.99%+ historically). Model inference errors returned in result field (not HTTP errors).

developers.cloudflare.com/workers-ai is comprehensive — model catalog with capabilities and pricing, REST API reference, Workers binding guide, and playground for testing models. Getting started: Cloudflare account (free), generate API token, first inference call in under 5 minutes. AI Playground in Cloudflare Dashboard for testing without code. Workers AI models page lists all available models with neuron cost estimates. Community via Cloudflare Discord (Workers AI channel) and Cloudflare Community Forums. Cloudflare engineering team active in community discussions.