APIRank
EN 中文

Cloudflare Workers AI

Listed at https://developers.cloudflare.com/workers-ai/

Overall Rank #5 ⭐⭐ Recommended
❌ Mainland China edge unreachable (CF has no in-China GPU POPs); proxy required; ideal for overseas / HK-MO-TW / outbound-China workloads | 🌍 International

💰 Token Pricing

TypePriceNote
Input Neurons-based: $0.011 per 1,000 Neurons (after free tier); per-model token prices vary per million tokens
Output Same Neurons billing; token price differs per model — Llama 3.3 70B $0.293/$2.253, GLM-5.3 $1.40/$4.40, DeepSeek V4 Pro $1.32/$3.96, GPT-OSS 120B $0.35/$0.75 per M tokens per million tokens
💡 Free Credits: 10,000 Neurons per day free on Workers Free plan

🤖 Supported Models (86)

@cf/zai-org/glm-5.3 — Z.ai 320B MoE coding model (1M ctx, $1.40/$4.40)@cf/zai-org/glm-5.3-flash — fast variant ($0.15/$0.50)@cf/deepseek-ai/deepseek-v4-pro-0813 — 1M-ctx reasoning ($1.32/$3.96)@cf/deepseek-ai/deepseek-v4-flash-0731 — fast reasoning ($0.44/$1.32)@cf/moonshotai/kimi-k2.7-code — code-tuned ($0.95/$4.00)@cf/openai/gpt-oss-120b — open-source GPT-OSS 120B ($0.35/$0.75)@cf/meta/llama-3.3-70b-instruct-fp8-fast — fast Llama ($0.293/$2.253)@cf/meta/llama-4-scout-17b-16e-instruct — Llama 4 ($0.270/$0.850)@cf/google/gemma-3-12b-it — Gemma 3 12B ($0.345/$0.556)@cf/mistralai/mistral-small-3.1-24b-instruct — Mistral Small 3.1 ($0.351/$0.555)@cf/ibm/granite-4.0 — IBM Granite 4.0 (text generation)@cf/black-forest-labs/flux-1-schnell — image generation@cf/openai/whisper — automatic speech recognition@cf/microsoft/resnet-50 — image classification

✨ Pros

  • ✓86-model catalog spanning LLM, embedding, image, and audio
  • ✓10,000 Neurons per day free, no credit card required
  • ✓Unified Neurons billing: $0.011 per 1,000 Neurons, with per-model token prices visible in the catalog
  • ✓Frontier-tier models at parity: GLM-5.3, DeepSeek V4, Kimi K2.7-code, GPT-OSS 120B
  • ✓Inference on Cloudflare's 300+ city edge, nearest GPU POP serves the request
  • ✓REST + Workers binding + OpenAI-compatible `/v1/chat/completions` — three entry points

⚠️ Cons

  • ×Workers Free plan excludes frontier models (GLM-5.3, DeepSeek V4, Kimi K2.6/2.7-code)
  • ×No in-China GPU POP; latency and reachability depend on international routing
  • ×Neurons billing is non-obvious for new users (must cross-reference token pricing)
  • ×Workers Paid requires a $5/month base subscription on top of usage
  • ×Model catalog is curated by Cloudflare — no BYO-model self-hosting

🎯 Best For

Globally-deployed AI applications; teams already running on Cloudflare Workers; startups needing frontier models with low operational overhead; agent, RAG, and edge inference workloads

💰 Pricing & Plans

TierDaily Free AllowanceNeuron RateNotes
Workers Free10,000 Neurons / day$0Generous daily free tier for prototyping; resets at 00:00 UTC; excludes frontier models (GLM-5.3, DeepSeek V4, Kimi K2.6/K2.7-code)
Workers Paid10,000 Neurons / day free + overage billed$0.011 per 1,000 NeuronsPay-as-you-go; per-model token prices displayed in catalog; frontier models accessible; AI Gateway credits unlock higher rate limits

🔧 API & Developer Experience

  • •Workers AI Binding: Call models directly from a Cloudflare Worker via `env.AI.run("@cf/...", { ... })` with no separate SDK or API key management — the Workers runtime handles auth and routing.
  • •REST API: Every model is also exposed via a plain HTTPS REST endpoint (https://api.cloudflare.com/client/v4/accounts/<id>/ai/run/<model>) so any non-Worker client can hit Workers AI without writing a Worker.
  • •OpenAI-Compatible: An OpenAI-shaped `/v1/chat/completions` endpoint is available (with API-token auth) so existing OpenAI SDK code works after a base-URL change.
  • •Streaming: Server-Sent Events streaming is supported on the Workers binding and REST endpoint, returning incremental tokens for low time-to-first-token UX.
  • •Vectorize Integration: Embeddings models output vectors that can be written directly to Cloudflare Vectorize (or any vector DB) from the same Worker — no intermediate HTTP hop.
  • •Edge Network: Inference runs on Cloudflare's global edge in 300+ cities; the closest GPU-bearing POP serves the request, giving sub-second TTFT for most regions.

🎯 Edge Serverless GPU & Frontier Model Catalog

Cloudflare Workers AI is not a hosted model API in the traditional sense — it is serverless GPU compute priced in Cloudflare's own 'Neurons' unit, accessible from any Worker or REST call. The August 2026 catalog spans 86 models across text, embedding, image, and audio, including frontier-tier models that other serverless platforms price as separate SKUs: Z.ai GLM-5.3 at $1.40/$4.40 per million tokens with 1M context (released 2026-08-28), DeepSeek V4 Pro at $1.32/$3.96 with 1M context, Moonshot Kimi K2.7-code at $0.95/$4.00 for code workloads, and OpenAI GPT-OSS 120B at $0.35/$0.75. Per-model granular pricing lets you route small classification jobs to Llama-3.2-1B at fractions of a cent while reserving frontier models for code or long-context tasks.

🌐 Regional Availability & Latency

Workers AI runs inference on Cloudflare's global edge in 300+ cities, so latency from any major region is governed by the nearest GPU-bearing POP rather than a single origin. For North America, Europe, East Asia, and Oceania the typical TTFT is sub-second on small models and 1-2 seconds on 70B+ frontier models. Mainland China is the explicit exception: Cloudflare's network does not currently serve inference traffic from in-China POPs, so mainland clients reach the global edge over the public Internet with variable latency (often 200-600 ms even on small models). Teams serving Chinese users either fall back to a domestic provider (Zhipu, Qwen, Doubao) or accept higher edge latency in exchange for Workers AI's unified billing and curated frontier catalog.