Cloudflare Sells 10,000 AI Agents a Second—and Bills You per Neuron

The pitch is the price sheet

Cloudflare’s Workers AI pricing page reads less like a cloud bill and more like a dare. Inference is sold in units of 1,000 neurons, and the free allocation is 10,000 neurons per day. A small model call costs a fraction of a neuron; a large one costs several. The company doesn’t hide the meter behind a “contact sales” wall.

That matters because most agent platforms price in tokens, which is a proxy for compute nobody can audit. Cloudflare prices in a unit it defines, publishes, and lets you watch tick down in the dashboard. Whether the neuron is a fair unit is a separate question. That it exists is the story.

What 10,000 agents actually cost

Run one agent that calls a model every few seconds and the free tier vanishes before lunch. Run 10,000 agents and the arithmetic gets uncomfortable fast. Cloudflare’s own developer docs put the per-neuron rate for popular text-generation models in the low fractions of a cent per thousand neurons, which sounds trivial until you multiply by a fleet.

The useful comparison is not against a frontier API. It is against the cost of keeping those agents alive: Durable Objects for state, Queues for handoffs, Workers for routing. Cloudflare sells all of it, and the AI line item is the smallest part of the bill. The company is not really competing on inference price. It is competing on the fact that the inference sits next to the state.

The catch is the unit you can’t convert

A neuron is not a token, and Cloudflare does not publish a fixed conversion. The same prompt can cost different neuron counts depending on model, batch size, and whether the request hits cache. Builders who budget in tokens will find their forecasts off by a factor they cannot predict from the price sheet alone.

That is not a scandal. It is a design choice: Cloudflare is billing for compute rather than for text, which is arguably more honest. But it means cost modeling requires measurement, not arithmetic. The teams that will do well here are the ones that log neuron consumption per agent action and treat it like any other resource metric.

Where the free tier is genuinely useful

The 10,000-neuron daily allowance is enough to run a small internal agent, a classification job, or a prototype that never leaves staging. It is not enough to run a production fleet, and Cloudflare does not pretend otherwise.

The smarter use is as a cost ceiling during development. If an agent burns the daily free allocation in an hour of testing, that is a signal about its loop behavior before it ever reaches a paid environment. Most agent cost blowups are loop bugs, not model prices.

What to watch

Cloudflare has been steadily pushing Workers AI toward agent workloads with Durable Objects, Vectorize, and AI Gateway. The pricing page is the least glamorous part of that stack and the one that will decide whether builders stay.

If the company publishes a neuron-to-token conversion table, the comparison shopping gets real. Until then, the honest advice is to instrument first and forecast second. The bill is legible. The unit is not yet.