Cloudflare Workers AI
Pay-per-use open models on Cloudflare's own GPUs, with AI Gateway for caching, rate limits and routing.
Overview
Workers AI runs open models on Cloudflare's own GPUs across its global network, billed per use. It suits teams already building on Cloudflare Workers or Pages. What we like: you start with a single Workers binding; prices are low, especially on small models (Llama 3.3 70B costs about US$0.293 per million input tokens and US$2.253 per million output tokens; gpt-oss-120b about US$0.35 / US$0.75); and AI Gateway's analytics, caching and rate limiting are free. Watch out for: you choose from Cloudflare's catalogue and can't upload your own weights yet (bring-your-own-model via Replicate's Cog is coming). Fine-tuning is LoRA only, on non-quantized base models (adapters under 300 MB, rank up to 32, in open beta). The largest models need a Workers Paid plan, and Unified Billing adds 5% to credit purchases. Pricing is per Neuron: US$0.011 per 1,000 Neurons, with 10,000 Neurons a day free. Cloudflare changes prices often, so check its pricing page for current rates.
Specifications
| Hosting type | Serverless API |
|---|---|
| Model families | Llama, Mistral, Qwen, gpt-oss |
| LLM models | Llama 3.1 8B Instruct, Llama 3.3 70B Instruct, Mistral Small 3.1, Qwen2.5-Coder 32B Instruct, Qwen3 30B-A3B, gpt-oss-120b, gpt-oss-20b |
| Pricing range | From US$0.045 |
| Pricing model | Pay per token |
| Free tier available | Yes |
| Data center locations | |
| Hardware / GPU options | Fully managed. Models run on Cloudflare's own GPUs in its data centres worldwide; you don't choose hardware. |
| VRAM & memory | |
| Maximum model size | |
| OpenAI-compatible API | Yes |