Cloudflare Workers AI

Learning Objectives

• Understand serverless GPU inference and where Workers AI fits the AI compute stack
• Identify the Neuron-based pricing model and the per-token rates for large models
• Evaluate when Workers AI beats a closed-model API or self-hosted GPU inference
• Understand the October 2026 GLM-5.3 cache exposure and what it teaches about shared inference

What Is Cloudflare Workers AI?

Workers AI is Cloudflare's serverless GPU inference platform. You call a model and pay for the inference, with no GPU rental, no instance management and no cold-start delays. It is the AI inference layer of the Cloudflare developer platform, alongside Workers (compute), R2 (storage), D1 (SQL), KV (key-value) and Durable Objects.

Where AWS Bedrock and Azure sell access to closed flagship models, Workers AI runs open-weight models on Cloudflare-managed GPUs, plus a handful of partner models. The catalog has changed character since launch. It began with small task models for classification, embeddings and transcription; by late 2026 it also hosts large frontier-class open models, including Zhipu's GLM-5.3 with a context window of about one million tokens, Moonshot's Kimi K2.6 and K2.7-Code, DeepSeek V4 Pro and V4.1 Flash, OpenAI's gpt-oss models, NVIDIA's Nemotron 3 and Meta's Llama 4 Scout. Requests are routed to available GPU capacity on Cloudflare's network, so latency depends on where that capacity sits rather than on a single cloud region.

Pricing

Workers AI pricing is Neuron-based. Neurons measure the GPU compute a request consumes, and every account gets 10,000 Neurons a day free, reset at midnight UTC. Above that, usage costs 1.1 cents per 1,000 Neurons on the Workers Paid plan, which has a minimum charge of five dollars a month. For large language models Cloudflare also publishes per-token rates, which are easier to compare with other providers. As of Cloudflare's October 2026 rate card, GLM-5.3 costs one dollar forty per million input tokens and four dollars forty per million output tokens, Kimi K2.6 costs 95 cents in and four dollars out, and gpt-oss-120b costs 35 cents in and 75 cents out. The newest large models, including the GLM-5 series, Kimi K2.6 and later, and DeepSeek V4, require Workers Paid or prepaid AI Gateway credits rather than the free allocation.

The free tier is genuinely usable for prototyping and small projects, since 10,000 Neurons a day covers thousands of small-model calls. It covers far fewer calls to the large models, which is why those sit behind the paid plan. For high-volume production traffic, compare the per-token rates with self-hosted GPU economics.

Core Features

Serverless GPU Execution

Pay only for the inference you run, with no idle GPU rental and no instance management. Cloudflare manages the GPU pool and routes each request to available capacity.

A Broader Model Catalog

Beyond the large language models, the catalog covers text embeddings, text classification, speech recognition, text-to-speech, text-to-image, image-to-text, image classification and translation. Partner models include Deepgram's Aura voices and Nova and Flux speech recognition, and Leonardo.AI's image models. Cloudflare also builds its own small "decision" models, the Clef family, which return probabilities for typed questions rather than chatting.

Tight Workers Integration

Call Workers AI from a Workers script with env.AI.run('model-name', { prompt: '...' }), with no API key juggling and no separate SDK. The integration is designed for developers building AI features inside web apps on the Cloudflare developer platform.

AI Gateway

AI Gateway sits in front of any model endpoint, including Workers AI, OpenAI and Anthropic, and adds caching, rate limiting, retries, analytics and a single observability view. It is the usual way to mix Workers AI with closed models from other providers.

Vectorize (Vector Database)

Cloudflare's vector database for embeddings pairs with Workers AI embedding models for retrieval-augmented generation (RAG) workloads, so vectors and inference run on the same network.

Stripe Projects — Agentic Onboarding (May 2026 Open Beta)

In May 2026, Cloudflare and Stripe opened a beta of a Stripe Projects integration that lets AI agents onboard to the Cloudflare developer platform themselves: an agent can create a Cloudflare account, buy a domain, and deploy a Worker without a human clicking through forms. Stripe handles identity checks, issues the agent a scoped payment token rather than a real card number, and enforces a default spending cap of 100 dollars a month per provider. A human still grants the initial permission and accepts the terms of service.

The GLM-5.3 Cache Exposure (October 2026 Disclosure)

On October 26, 2026, Cloudflare disclosed that between August 27 and September 9, 2026 its hosted GLM-5.3 model could return fragments of other customers' recent requests. Two problems overlapped. The model was configured to accept about one million tokens of context, but the memory that holds a request's working state on the first stage of servers fit only about 681,000 tokens. And a bug in SGLang, the open-source serving software, marked over-long requests as rejected without stopping them, so the second stage read leftover memory from earlier requests. Cloudflare counted 155 affected responses, all to very long prompts, and found no sign of deliberate exploitation after a researcher reported the problem on September 8. It took the model offline for several hours on September 9, enlarged the memory pool, patched SGLang upstream, applied the fix to every model it serves with SGLang, and told other operators to upgrade to SGLang 0.5.22 or later. No other Workers AI model was affected, and Cloudflare's post did not say whether the affected customers were contacted individually.

Strengths

• Generous free tier: 10,000 Neurons a day with no credit card, meaningful for prototyping
• Large open models without GPUs: GLM-5.3, Kimi K2.6, DeepSeek V4 and gpt-oss are available per token, with no hardware to run
• Published per-token rates: large-model prices convert cleanly for comparison with other providers, and cached input is discounted on several models
• Cloudflare ecosystem fit: tight integration with Workers, R2, D1, Vectorize and AI Gateway, a single platform for full-stack AI apps
• No GPU operations: Cloudflare manages the GPU pool, and you write code
• Transparent incident handling: the GLM-5.3 disclosure named the root cause, the exposure window, the response count and the upstream fixes

Limitations & Considerations

• Shared inference can leak between customers: the 2026 GLM-5.3 exposure returned fragments of other customers' requests in 155 responses over two weeks; keep secrets out of prompts where you can, and weigh this risk for regulated data
• No frontier closed models: the leading closed models from OpenAI, Anthropic and Google are not on Workers AI; reach them through AI Gateway instead
• Licenses vary by model: open weights do not mean identical terms, and several catalog models carry community or modified licenses, so check each model's license before building a commercial product on it
• The largest models need the paid plan: the free allocation does not cover the GLM-5 series, the newer Kimi models or DeepSeek V4
• Cost at extreme scale: for very high-volume inference, self-hosted GPU economics may beat per-request pricing, so model your specific workload
• Lock-in risk: AI Gateway, Workers AI and Vectorize are tightly coupled, so moving off Cloudflare means rewriting the AI stack
• Cloudflare dependency: outages on Cloudflare's network affect Workers AI alongside the rest of the developer platform

Best Use Cases

When to choose alternatives:

• Closed frontier models needed → OpenAI API, Anthropic API or Google Gemini API, with AI Gateway in front of them
• Massive training workloads → a dedicated GPU cloud such as Lambda Cloud or CoreWeave, or a hyperscaler AI service
• Custom-trained model hosting → Modal, Replicate or AWS SageMaker for arbitrary container deployment
• Strict data isolation → a dedicated endpoint or self-hosted deployment, so no other customer shares your serving stack

Key Takeaways

• Cloudflare Workers AI is serverless GPU inference for open-weight models, now including large frontier-class models such as GLM-5.3, Kimi K2.6 and DeepSeek V4 as well as small task models
• Pricing is Neuron-based, with 10,000 Neurons a day free and 1.1 cents per 1,000 Neurons after that; the largest models are priced per token and need the Workers Paid plan
• In October 2026 Cloudflare disclosed that its GLM-5.3 service leaked fragments of other customers' requests in 155 responses between August 27 and September 9, caused by a context-length misconfiguration and an SGLang bug that it has since fixed upstream
• Pair it with AI Gateway to mix open models on Workers AI with closed models from OpenAI, Anthropic and Google
• Check each model's license before commercial use, since open weights do not all come with the same terms
• The May 2026 Stripe Projects beta lets AI agents create Cloudflare accounts, buy domains and deploy Workers, with Stripe handling identity checks, scoped payment tokens and a default 100 dollars a month spending cap