Cloudflare Workers AI

Cloudflare Workers AI runs open models on Cloudflare's network and bills by the token, through Workers bindings or a REST API. On 31 October 2026 gpt-oss-120b cost $0.35 per million input tokens and $0.75 per million output, and every account gets 10,000 neurons a day free. It is a cheap, fast start for light and medium use, with less control than renting a GPU.

Some links on this page go through /go/ and may earn us a fee; it never changes our order.

Facts at a glance

To compare it with GPU clouds and other APIs, use the hosting compare and read the AI hosting overview.

How it works

Workers AI lives next to Cloudflare Workers, the platform that runs your code at the edge. From a Worker you bind to an AI service and call a model by name. From outside, you call a REST endpoint with an API token. Cloudflare documents OpenAI-compatible endpoints too, so existing OpenAI client code works with a different base URL and model name, and that makes switching in or out cheap.

Billing is in neurons, which the pricing page describes as Cloudflare's way of measuring AI output, representing the GPU compute needed for a request. The page says it has made pricing more granular by showing per-model unit prices, such as per million tokens, while still billing in neurons on the back end. Each model row shows both units, so you can compare them with other providers.

The prices

From the Workers AI pricing page on 31 October 2026, in dollars per million tokens:

Cached input is cheaper on the models that list it, for example $0.014 per million on DeepSeek V4 Flash 0731. The page also lists embeddings, image, audio and other model types, with a Qwen3 embedding at $0.012 per million input tokens.

Plans and the free allowance:

A few models, including Kimi K2.6, the GLM 5 series and the DeepSeek V4 models, require a paid billing method, either Workers Paid or prepaid AI Gateway credits. Credits can pay for Workers AI inference through unified billing, and requests to frontier models that use prepaid credits get higher rate limits, per the page.

What you pay in practice

The volumes are examples we chose: a team that sends 40 million input tokens and 8 million output tokens a month.

Other examples. A prototype that uses 2 million tokens a month on a small model may sit inside the free allowance, which is worth up to 300,000 neurons in a 30-day month. At $0.011 per 1,000 neurons that is $3.30 a month of free compute. A coding agent loop that burns 500 million input tokens a month on gpt-oss-120b would cost $175 for input alone, and output on top. At that volume, price a rented GPU. A Runpod Secure A100 at $1.79 an hour costs about $1,289 for a full month, so the API stays cheaper unless volume is several times higher, or the data must stay under your own control.

Add the cost of the Worker that calls the model. The Workers pricing page lists a Free plan with 100,000 requests a day and 10 ms of CPU per request, and a Paid plan from $5 a month with 10 million requests and 30 million CPU milliseconds included, then $0.30 per extra million requests, with no egress charges. That is small next to the token bill.

Limits

Fine-tuned inference with LoRA adapters is available on selected base models: you upload an adapter and run it. Check the docs for current model support and size limits.

Data and privacy

Cloudflare's data page says it processes certain customer data to provide Workers AI. The page lists your inputs, outputs, embeddings and any training or adaptation data you supply as Customer Content. On retention, it says Workers AI processes Customer Content only for as long as needed to provide the service, and that for standard inference requests that do not invoke a stateful feature, Cloudflare does not retain inputs or outputs after they are no longer needed to process the request and deliver the response, subject to limited exceptions.

Two cautions follow from the page itself. If you route requests through AI Gateway to another AI provider, that provider may process and retain the content under its own terms. And stateful features, such as fine-tuned adapters, store data by design. For regulated data, get the Data Processing Addendum and have it read.

Regions

Workers AI is part of Cloudflare's global network, so requests are served from locations near the user or from where capacity exists. . If region pinning is a requirement, check that first, or choose a host that lets you pin, such as Modal or Hugging Face Inference endpoints.

The neuron arithmetic

Because billing runs on neurons, it helps to know how the two units relate. The pricing page shows Llama 3.1 8B fp8 fast at $0.045 per million input tokens, which it states as 4,119 neurons, and at $0.384 per million output tokens, which is 34,868 neurons. Divide dollars by neurons and you get the same rate as the headline: $0.045 over 4,119 is about $0.0000109 per neuron, close to $0.011 per 1,000.

That gives you a quick way to read the free allowance in tokens. The 10,000 daily neurons buy about 2.4 million input tokens a day on Llama 3.1 8B (10,000 divided by 4,119, times a million), or about 287,000 output tokens. On the 70B model, which lists 26,668 neurons per million input tokens, the same allowance covers about 375,000 input tokens. So the free tier is generous on small models and brief on big ones. These conversions are our arithmetic from the page's figures.

Output costs far more than input on most models, as in the usual token price pattern. Llama 3.3 70B's output is $2.253 against $0.293 for input, a ratio of almost eight to one. If your agent generates long reasoning traces, output dominates the bill. If it reads large files and answers briefly, input does.

Workers AI against other token APIs

At the same listed model, Groq and Ollama's cloud undercut Cloudflare's rate, and Cerebras matches it. Cloudflare's case rests on integration: the same account, the same bill, bindings from Workers, and a free allowance. If you are not on Cloudflare, the integration argument weakens and price becomes the main test. See Groq vs Cerebras for the speed side.

Building an agent on Cloudflare

A typical stack puts the agent's code in a Worker, state in a store such as D1 or Durable Objects, and the model on Workers AI. This site's own plan runs on Cloudflare, so we have a bias to disclose. Neither the Workers nor the Workers AI pages we read say a Worker will outlast a long agent loop, so check the current limits for wall-clock time and subrequests before you design a long-running agent on it.

A few design habits help whichever host you pick:

1. Keep prompts short. Input tokens are cheap, but a long system prompt on every call adds up. Cached input is cheaper on models that list it.
2. Cap output. Set a maximum output length per call. Output tokens cost most.
3. Cache answers. Identical questions need not hit the model twice.
4. Use a small model for easy steps. Classification and routing rarely need a 70B model. Llama 3.2 1B lists $0.027 and $0.201 per million.
5. Log token counts per user. It is how you spot the one customer who costs you more than they pay.

For the wider pattern, see how to build an AI agent.

Moving off Workers AI

Because the API is OpenAI-compatible, moving is a change of base URL, key and model name. Keep your prompts and evaluation set in your own repository. If you used LoRA adapters, keep the adapter files and training data, since the adapters are tied to Cloudflare's base models. If you need a GPU of your own later, the Runpod and Lambda Cloud pages show the rented route, and vLLM is the usual server.

Who should pick Workers AI

Alternatives

• Hugging Face Inference: dedicated endpoints and provider routing.
• Replicate: per-second model hosting, now joining Cloudflare per its blog.
• Modal: your own code on a GPU.
• Runpod and Lambda Cloud: rented GPUs.
• AWS Bedrock: models inside AWS.
• Groq and Cerebras: see Groq vs Cerebras. Groq lists gpt-oss-120b at $0.15 and $0.60 per million tokens, below Cloudflare's rate.
• Gateways: OpenRouter alternatives.

Cloudflare also has an MCP server for its platform, covered at Cloudflare MCP.

Getting started

1. Create a Cloudflare account and an API token with Workers AI permission.
2. Call a model from a Worker with the AI binding, or from the REST endpoint.
3. For existing OpenAI code, change the base URL, the key and the model name.
4. Watch the neuron count in the dashboard against the 10,000 free.
5. If you need paid-only models, move to Workers Paid or add AI Gateway credits.

Check the exact request format against the current docs, since model names and fields change.

Our take

Workers AI is the cheapest low-effort way to call an open model if you already use Cloudflare, and the daily free allowance makes experiments free. For gpt-oss-120b, Groq's listed rate is lower, so compare before you pick on price alone. The data page is clear for standard requests, though the exceptions deserve a read. The weak points are model choice, regional control and the absence of dedicated capacity. Our conclusions come from Cloudflare's pricing, limits and data pages, and we ran no benchmarks.

Related pages

• Hosting hub and AI hosting overview
• Private AI hub
• Cloud vs self-hosted AI