Cloudflare Workers AI gives 10,000 free Neurons daily then charges $0.011 per 1,000

A post on dev.to breaks down what Cloudflare Workers AI actually costs as of October 2026, and the numbers are concrete enough to budget against: every account receives 10,000 Neurons per day at no charge, and usage beyond that is billed at $0.011 per 1,000 Neurons on the Workers Paid plan.

The daily allowance and the paid tier

According to the post, which draws on Cloudflare's documentation, the 10,000-Neuron allowance resets at 00:00 UTC each day and applies to both Workers Free and Workers Paid accounts. Over a 30-day month that works out to 300,000 Neurons at no cost. Spending past the allowance requires upgrading to Workers Paid, a plan with a $5 monthly minimum, though some models also accept prepaid AI Gateway credits as an alternative.

The reset cadence has a practical consequence the author points out: a batch job that runs across midnight UTC draws on two separate daily allowances. For larger workloads the arithmetic stays simple, since one million Neurons above the free tier comes to 1,000 times $0.011, or $11.

Model choice dominates the token bill

Behind the flat Neuron meter, per-token prices diverge sharply depending on the model. Input and output are priced separately per million tokens, and the post lists four examples:

Model	Input per million tokens	Output per million tokens
@cf/meta/llama-3.2-1b-instruct	$0.027	$0.201
@cf/meta/llama-3.1-8b-instruct-fp8-fast	$0.045	$0.384
@cf/openai/gpt-oss-120b	$0.350	$0.750
@cf/deepseek-ai/deepseek-v4-pro-0813	$1.320	$3.960

The spread between the cheapest and priciest input price in that set is roughly 49x. The author's advice is to run real prompts against the smallest model that produces acceptable answers on your own test set before considering anything larger.

Rate limits arrive before the bill

Throughput, not money, is often the first constraint. Text generation allows 300 requests per minute, but models that require the Workers Paid plan are capped at 20 requests per minute per account and per model on standard billing, or 50 with prepaid AI Gateway credits under unified billing. Other task types carry their own ceilings: 3,000 requests per minute for text embeddings (1,500 for @cf/baai/bge-large-en-v1.5) and 720 for translation. The author gives a concrete example of the mismatch: a queue consumer that fans out 100 text-generation calls in a minute to a paid-only model will be throttled at 20 or 50, no matter how much budget sits behind it.

Calling a model from a Worker

Getting started requires little setup. You add an ai binding to the Wrangler configuration, then call env.AI.run from the Worker's fetch handler with a model identifier and a messages array, and deploy with npx wrangler deploy. One caveat worth knowing before testing: npx wrangler dev runs locally but still executes against your Cloudflare account, so local development consumes the free allocation.

Why it matters

Serverless inference pricing is frequently opaque, with per-token rates that are hard to translate into a monthly figure. Cloudflare's Neuron unit, a fixed daily allowance and a published per-1,000 rate give developers an anchor for planning: 10,000 free Neurons a day is enough for meaningful prototyping, and the first million Neurons past that cost $11.

Just as important, the post separates cost from throughput. A team can budget correctly and still watch jobs get throttled at 20 requests per minute, so capacity planning for paid-only models has to start from the rate limit rather than the invoice. Two caveats the author flags: prices can change, so check Cloudflare's pricing page before committing a budget, and the documentation does not publish a Neuron-to-token conversion, meaning per-call cost estimates still require logging your own usage.