Cloudflare Workers AI
6.8/10Serverless GPU inference service from Cloudflare that runs a curated catalog of open-weight AI models on its global network, billed per use.
Data last updated November 2026 · Our methodology
Overview
Cloudflare Workers AI lets developers run inference on open-weight models without provisioning GPUs. The catalog covers text generation (Llama, Mistral, Qwen, DeepSeek, gpt-oss), embeddings, Whisper speech-to-text, image generation, vision, and translation. Models are called from a Worker through an AI binding or from any other stack through a REST API, and OpenAI-compatible endpoints let existing SDK code switch over by changing the base URL. It fits best for teams that already build on Cloudflare Workers, Pages, and Vectorize and want edge classification, embeddings for retrieval, summarization or translation middleware, or cheap prototypes. Pricing is measured in Neurons. Every account gets 10,000 Neurons a day free, and usage above that costs $0.011 per 1,000 Neurons on the Workers Paid plan, which reviewers put at $5 a month. Per-model token prices are published, ranging from $0.027 per million input tokens for small Llama models to $1.32 input and $3.96 output per million tokens for DeepSeek V4 Pro. AI Gateway can sit in front of it to add caching, rate limiting, retries, and request analytics. Its main limitation is that you can only run models from Cloudflare's catalog. There are no proprietary frontier models and you cannot deploy custom architectures, though LoRA adapters are supported in beta. Rate limits are set per model, and premium text models are capped at 20 requests per minute per account under standard billing.
Our Rating
AI Software6.8/10
Hosting & Infrastructure6.7/10
Who Cloudflare Workers AI is best for
Best software for DevelopersGood fit
You reach it through code, using Worker bindings, a REST API, and OpenAI-compatible endpoints.
The daily free Neurons and low per-token prices make prototyping cheap.
It suits embedding classification, embeddings, and summarization in products, but per-model rate limits need planning.
Less suited
Enterprises can run it on Cloudflare's network, but the catalog-only models and RPM caps limit heavy production use.
Pros & Cons
Pros
- 10,000 free Neurons a day plus low per-token rates, such as $0.045 per million input tokens for Llama 3.1 8B fast, keep small workloads cheap.
- OpenAI-compatible endpoints mean existing SDK code needs little more than a URL change.
- Runs next to Workers, Vectorize, and AI Gateway, which shortens the path to building edge RAG apps.
- No GPU provisioning or idle-capacity billing.
Cons
- Only catalog open-weight models are offered, with no proprietary frontier models and no custom architectures.
- Premium text-generation models are capped at 20 RPM per account (50 RPM with prepaid AI Gateway credits).
- Reviewers report variable cold-start latency and smaller context windows than frontier APIs.
- Neurons do not map cleanly to tokens, which makes cost forecasting harder.
- LoRA fine-tune support is in beta and limited to some models.
Key Features
Cloudflare Workers AI Pricing
Workers Free
- Resets at 00:00 UTC
- 10,000 Neurons per day
- No access to premium paid-only models
- Requests fail once the daily limit is reached
Workers Paid
- 10,000 free Neurons per day
- Pay-as-you-go above the allocation at $0.011 per 1,000 Neurons
- Access to premium models such as Kimi K2.6, GLM 5.2, and DeepSeek V4 Pro
Frequently Asked Questions
Yes. Every account gets 10,000 Neurons per day, reset at 00:00 UTC. On the free plan, requests fail once you pass that limit.
On Workers Paid, usage above the daily allocation costs $0.011 per 1,000 Neurons. Cloudflare also publishes per-model token prices.
Not natively. Only Cloudflare's catalog of open-weight models runs on Workers AI. AI Gateway can route requests to other providers.
Limits depend on the task type. For example, text generation allows 300 RPM and embeddings allow 3,000 RPM. Paid-only text models are limited to 20 RPM per account.
A review summarizing Cloudflare's documentation says Cloudflare does not use customer content to train models without consent.
LoRA adapters are supported for selected models, and that support is still in beta.
Similar Tools
See all alternatives
Amazon's public cloud platform, with over 200 pay-as-you-go services for compute, storage, databases, networking and AI, used by startups through large enterprises and governments.

Vercel Sandbox runs untrusted or AI agent-generated code in isolated Firecracker Linux microVMs on demand. It is aimed at developers building coding agents, code playgrounds and preview tools.

Baseten is an AI inference platform that lets engineering teams serve, scale and train open-weight and custom models on managed multi-cloud GPUs, billed per minute.
Ranked near Cloudflare Workers AI in AI Software
See Cloudflare Workers AI in the listCloudflare Workers AI is #136 in AI Software by our score.




