Cloudflare Workers AI Free Tier: 10,000 Neurons/Day (2026)

Short answer: every Cloudflare account can run open models on Workers AI for free, up to 10,000 Neurons per day, resetting at 00:00 UTC. That is a compute budget, not a request count, so how far it goes depends on which model you call. On the free Workers plan the budget is a hard cap; on Workers Paid ($5/month minimum) anything above it costs $0.011 per 1,000 Neurons.

What you actually get

Workers AI is Cloudflare's serverless GPU inference service. You don't rent a GPU or deploy a model; you call a model from Cloudflare's catalog (text generation, embeddings, image generation, speech-to-text, translation, classification and more) by name, and Cloudflare runs it on GPUs in its network. The free part is a daily allocation of compute measured in Neurons.

In dollar terms the daily allocation is worth 10,000 ÷ 1,000 × $0.011 = $0.11 per day, or roughly $3.30 over a 30-day month. That sounds small, but many of the models in the catalog are small, cheap open models, so the allowance covers a surprising amount of prototyping traffic.

How Neurons translate into real requests

Cloudflare's pricing page lists each model's price per million tokens alongside the equivalent number of Neurons. The figures below are taken from that page; the request estimates are our own arithmetic, assuming a typical chat request of 500 input tokens and 300 output tokens.

Worked example for Llama 3.1 8B: 500 input tokens × 25,608 ÷ 1,000,000 = 12.8 Neurons, plus 300 output tokens × 75,147 ÷ 1,000,000 = 22.5 Neurons, for about 35.3 Neurons per request. 10,000 ÷ 35.3 ≈ 283 requests per day. Output tokens usually dominate, so asking for shorter answers is the single easiest way to stretch the budget.

Embeddings and images

• Embeddings are very cheap. @cf/baai/bge-m3 and @cf/qwen/qwen3-embedding-0.6b cost 1,075 Neurons per million input tokens, so the free 10,000 Neurons cover roughly 9.3 million tokens of embedding input per day — enough to index a sizeable documentation site for a RAG prototype.
• bge-small-en-v1.5 is 1,841 Neurons per million tokens (about 5.4 million tokens/day free); bge-large-en-v1.5 is 18,582 (about 540,000 tokens/day).
• FLUX.1 schnell images are priced per 512×512 tile (4.80 Neurons) and per diffusion step (9.60 Neurons). A 1024×1024 image is 4 tiles; with 4 steps that is 4 × 4.80 + 4 × 9.60 = 57.6 Neurons, so about 170 images a day fit in the free allocation.
• Premium image models cost much more: Leonardo Lucid Origin is 636 Neurons per tile, so a single 1024×1024 image uses over 2,500 Neurons before step charges — only a few per day on the free allowance.

The 10,000 Neurons are one shared pool for the whole account, not a per-model allowance. Mixing a large model with embeddings and images all draws from the same daily budget.

Who qualifies

Anyone with a Cloudflare account. There's no application, student status, startup program or trial period: the allocation is part of both the Workers Free and Workers Paid plans and renews every day for as long as the account exists. Cloudflare doesn't publish country restrictions specific to Workers AI beyond its normal terms of service.

The one real eligibility split is by model. Cloudflare's pricing page says Kimi K2.6, Kimi K2.7 Code, GLM-5.2, GLM-5.3, GLM-5.3 Flash, DeepSeek V4 Flash and DeepSeek V4 Pro require either the Workers Paid plan or prepaid AI Gateway credits. If you're on Workers Free, stick to the rest of the catalog (Llama, Qwen, gpt-oss, GLM-4.7 Flash, embeddings, FLUX and so on).

Step-by-step: your first free Workers AI call

1. Create an account at dash.cloudflare.com/sign-up and verify your email. You start on the Workers Free plan.
2. In the dashboard, open Workers AI and select Use REST API.
3. Select Create a Workers AI API Token, review the prefilled permissions, then Create API Token. Copy it somewhere safe. (A custom token works too, but it needs both Workers AI read and edit permissions.)
4. On the same page, copy your Account ID under Get Account ID.
5. Send a request to a model, replacing the account ID and token (example below).
6. Optional: inside a Worker, add an AI binding in your Wrangler configuration and call env.AI.run() instead of the REST API — no token handling needed.

curl https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/ai/run/@cf/meta/llama-3.1-8b-instruct \
  -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \
  -d '{ "prompt": "Where did the phrase Hello World come from?" }'

Using the OpenAI SDK instead

Cloudflare also exposes an OpenAI-compatible base URL: https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1. It supports Chat Completions for most text models, Embeddings, and the Responses API for @cf/openai/gpt-oss-120b and @cf/openai/gpt-oss-20b (non-streaming only). Point any OpenAI client at that URL, pass the Workers AI token as the API key and use the @cf/… model name.

from openai import OpenAI

client = OpenAI(
    base_url=f"https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/v1",
    api_key=CLOUDFLARE_API_TOKEN,
)
resp = client.chat.completions.create(
    model="@cf/meta/llama-3.1-8b-instruct",
    messages=[{"role": "user", "content": "Give me three names for a cat."}],
    max_tokens=100,
)
print(resp.choices[0].message.content)

Rate limits (separate from the Neuron budget)

Even with Neurons left, requests are throttled per task type. These are Cloudflare's published per-minute limits; beta models may be lower, and inferences you run locally with Wrangler count toward the same limits.

Cloudflare accepts requests for higher limits or custom models through a Custom Requirements Form linked from the limits page.

What happens when you run out

• Workers Free: once the account has used 10,000 Neurons since the last 00:00 UTC reset, further Workers AI requests fail with an error. Nothing is billed; you simply wait for the reset or upgrade.
• Workers Paid: the first 10,000 Neurons each day are still free; usage above that is metered at $0.011 per 1,000 Neurons and added to your Cloudflare invoice. Workers Paid itself has a $5 per month account minimum, which also raises your general Workers limits (10 million requests per month included instead of 100,000 per day).
• Paid-only models: these need Workers Paid or prepaid AI Gateway credits no matter how many free Neurons you have left.

On Workers Paid there is no built-in hard cap at the free allowance. If a bug loops on a large model, overage keeps accruing at $0.011 per 1,000 Neurons. Add your own request counting or use AI Gateway rate limiting before exposing an endpoint publicly.

Best uses for the free allowance

• RAG prototypes: generate embeddings with bge-m3 (millions of tokens a day fit in the allowance), store them in Vectorize, and answer with a small instruct model.
• Edge features in an existing Worker: classify, summarise, translate or moderate user input inside the Worker that already serves your site, without a separate API provider or key.
• Thumbnail and placeholder images: FLUX.1 schnell at around 58 Neurons per 1024×1024 image leaves room for over a hundred images a day.
• Hobby bots and internal tools: a Discord or Slack bot on Llama 3.1 8B or Qwen3 30B-A3B comfortably fits a few hundred messages a day.
• Benchmarking open models: the same API and billing unit across dozens of models makes side-by-side comparisons cheap.

How it compares with other free AI APIs

Workers AI is unusual in giving a daily compute budget that never expires, rather than a one-time credit or a per-model request quota. That makes it a good permanent home for low-volume features, but its model list is mostly open-weight models.

If you're building a startup on Cloudflare, also look at Cloudflare for Startups, which offers platform credits that can cover usage beyond the daily allowance. For a broader overview of no-cost keys, see our guide to free AI API keys.

History and recent changes

• Before April 2024: Workers AI ran as an open beta.
• April 2, 2024: Cloudflare announced general availability. The GA post introduced the 10,000 free Neurons per day for every customer, a hard limit for Workers Free after that, and $0.011 per 1,000 additional Neurons on Workers Paid. It also raised most LLM rate limits from 50 to 300 requests per minute.
• Since then: the pricing page moved to showing per-model prices in dollars per million tokens next to the Neuron equivalents, while billing still happens in Neurons. Newer frontier models (Kimi, GLM-5.x, DeepSeek V4) were added as paid-plan-only, with their own lower rate limits.

Common mistakes

• Assuming 10,000 means 10,000 requests. It's 10,000 Neurons; a single large-model request can use hundreds.
• Forgetting the reset is in UTC. Depending on your time zone, the budget may refresh in the middle of your working day.
• Picking a paid-only model on Workers Free. Those calls fail regardless of your remaining allowance.
• Leaving max_tokens unset. Output tokens cost more Neurons than input on almost every model; cap them.
• Testing locally and being surprised by usage. Wrangler local mode still runs on Cloudflare GPUs and still counts.