Cloudflare Workflows step billing: your sleeps now cost money

Durable workflow engines sell you a comforting lie: that waiting is free. Cloudflare just put a price on it. As of August 10, Cloudflare Workflows pricing bills steps and storage on top of the requests and CPU time you were already paying — and a step.sleep() that idles for three days counts as a step the same as one that does real work.

What actually happened

Cloudflare announced the change on July 7 with a deliberate runway: no step or storage billing before August 10, 2026, so teams could look at their usage first. That date has now passed.

The pricing page gives the numbers. On Workers Paid: 500,000 steps included per month, then $0.80 per additional 100,000 steps; 1 GB-month of storage included, then $0.20 per additional GB-month. On Workers Free: 3,000 steps per day and 1 GB-month of storage, with no charges beyond the allowance. Requests and CPU time keep their existing Workers rates — 10 million requests and 30 million CPU milliseconds included on Paid.

Two definitions carry the whole story:

• A step is any unit of work the Workflow executes, including sleeping and waiting for events. Durable waits are billable units, not free parking.
• Retries and rollback handlers are excluded from the step count. So a flaky third-party API that forces five retries costs you CPU and requests, but not five steps.

That second one is more generous than most people assume, and it changes where you should look for waste.

Why serverless step pricing matters for your business

Your cost model moved from "how long" to "how many." Idle Workflows accrue no CPU charges — a workflow parked on sleep or waiting for a webhook is not burning compute. What it now accrues is a step, plus persisted state sitting in storage the entire time. A 30-day dunning sequence with a wait between each of eight reminders is eight steps and a month of storage, regardless of how little thinking it did.

Count steps per business event, not per month. 500,000 steps sounds enormous until you write it as a rate. An order-fulfillment workflow with 12 steps clears 500k at about 41,000 orders. An agent loop that does a step per tool call clears it a lot faster than that. Do the division once, in a spreadsheet, before the invoice does it for you.

The cheapest optimization is merging steps, not removing them. Steps exist so that work is durable across failures — each one is a checkpoint. Collapsing three sequential API calls that always succeed or fail together into one step keeps the same reliability and bills one third as much. Collapsing steps that genuinely need independent retry boundaries just moves the cost to CPU and makes failures worse. Know which is which.

And check what your state weighs. Storage is persisted Workflow state in GB-months. If you're passing whole API responses between steps because it was convenient, you are now renting that convenience by the month. Pass identifiers, fetch on demand.

This is the pattern we keep writing about from a different angle every few weeks: platform pricing is a design input, not an afterthought. A workflow shaped for a free tier and a workflow shaped for per-step billing are different workflows. Better to shape it once, knowingly.

Sources: Cloudflare changelog, Cloudflare Workflows pricing.