What changed
Cloudflare said step and storage billing for Workflows would begin no earlier than 10 August 2026. Its current pricing documentation now defines four billable dimensions for Workflows: requests, CPU time, storage and steps.
On Workers Paid plans, 500,000 Workflow steps are included each month, with additional steps priced at $0.80 per 100,000. One GB-month of persisted Workflow state is included, with additional storage priced at $0.20 per GB-month. Requests and CPU time continue to use the wider Workers pricing model. Free-plan users receive 3,000 steps per day and 1 GB-month of storage without paid overages.
Why it matters
Cloudflare Workflows is attractive precisely because it hides much of the machinery needed for durable execution: retries, waiting, persisted state and long-running multi-step jobs. That is especially useful for AI products, automation and human-in-the-loop systems where a task may pause for minutes or days.
The new pricing does not make Workflows expensive by default, but it changes what builders need to measure. A design with lots of tiny steps, long retention and large persisted results can cost differently from one that performs the same business task with fewer checkpoints and externalizes bulky state. The architecture of the workflow is now part of its unit economics.
A step is now an economic unit
Cloudflare bills each Workflow step as a unit of work, with the paid plan including 500,000 steps per month before an $0.80 charge for each additional 100,000. The pricing page notes that rollback handlers and retries are not counted as steps for billing.
For a low-volume internal automation this may barely register. At product scale, however, a workflow that expands one customer action into dozens or hundreds of persisted steps can turn step count into a meaningful design variable. Teams should measure completed-business-task cost rather than looking at the headline price of an individual step.
Persisted state also has a carrying cost
Workflows automatically persists state so executions can survive waits, failures and restarts. On the paid plan, the first 1 GB-month is included and additional storage costs $0.20 per GB-month.
Completed state can be retained for up to 30 days on the paid plan, and developers can configure a shorter retention period when they do not need that history. Large binary outputs still count toward storage, so Cloudflare recommends storing very large or long-lived artifacts elsewhere, such as R2, and persisting a reference instead.
Waiting is still cheap in the way that matters
A Workflow that is waiting on an API call, sleeping or otherwise idle does not consume CPU time. Waiting instances also do not count toward active concurrency limits. That preserves one of durable execution's main advantages: a process can pause for an approval, retry window or external event without paying for a continuously running server process.
The pricing change therefore does not turn Workflows into metered wall-clock compute. It adds charges around the durable structure wrapped around the compute.
AI automation makes this worth watching
Agentic and AI-heavy applications often create exactly the kind of workloads Workflows is designed for: multi-step calls, long waits, retries, external APIs and human approvals. Those applications can also generate unusually large intermediate outputs and unpredictable execution paths.
Builders using Workflows for agents should put step count and persisted-state size next to token spend in their observability. An AI task that is cheap at the model layer can still be inefficient if the surrounding orchestration repeatedly checkpoints or retains bulky results.