When a Claude Code Routine is worth it (and when it isn't)

By Daniel S. · June 24, 2026

The first question people ask about Claude Code's Routines — the research-preview feature that runs a saved prompt on a schedule, API call, or GitHub event from Anthropic-managed cloud infrastructure — is usually "is the cloud version more expensive than just running it locally?"

It isn't. The docs are direct: "Routines draw down subscription usage the same way interactive sessions do." Same model, same per-token math, no cloud-execution surcharge. What changes is the volume you'll burn through — and whether the work was a good fit for unattended automation in the first place. That's the question worth asking before you set one up.

The cost picture in two lines

Per token, parity. The interesting variable is the prompt cache. By default it has a 5-minute lifetime; a separate 1-hour tier exists for an extra cost. Cache writes are 1.25× input price at 5 minutes and 2× at 1 hour, but cache reads are 0.1× — a tenth of the base input rate.

A Routine scheduled daily blows past either window every run, so it pays full uncached input price for its whole context each time. A local session you're actively working in stays inside the 5-minute window and keeps hitting the cache for cents on the dollar. That's the entire cost asymmetry — not a different price list, just a different reuse profile.

The three conditions that make a Routine pay off

A Routine earns its keep when all three are true:

1. The work has to happen on a cadence, not on demand. Nightly backlog grooming, weekly docs-drift scans, a PR-review pass on every opened pull request. If you only want it run when you ask, an interactive session is cheaper and faster — you skip the cold-cache penalty and you get to course-correct mid-run.
2. The prompt can be self-contained. Routines run autonomously as full cloud sessions with no permission prompts and no chance for you to clarify. The docs put it plainly: "the prompt must be self-contained and explicit about what to do and what success looks like." If the work needs judgment calls you can't pre-specify, the autonomy itself becomes the cost — the model over-fetches and over-verifies to be safe, and your token bill reflects that.
3. It needs to run while you're away. The whole point is that the laptop can be closed. If you're at the keyboard anyway, you're paying for cloud orchestration you don't need.

When it isn't worth it

Inverting the three above: irregular, exploratory, or interactive work. A research thread where you're still figuring out what "done" looks like is the worst fit — you can't write a self-contained prompt for a question you haven't fully formed yet, and you'll burn tokens on a cloud session producing output you'd have steered differently in person.

The other anti-pattern is low-signal recurring runs. A daily routine on a feed that only has news twice a week still fires seven times — you pay for five low-signal sessions to catch two real ones. Either widen the cadence (the minimum is hourly, but daily and weekly presets exist) or trigger it from an actual event instead. Routines support API triggers and GitHub triggers precisely so you can fire on signal rather than on clock.

The cache + cadence trick

If a Routine genuinely needs to run often, you can claw back some of the cache cost: opt the prompt into the 1-hour cache tier and pick an hourly schedule preset. Writes cost 2× input the first time, then every run inside that window pays 0.1× for the cached portion. For a tight, repeated prompt that's a real saving. It doesn't help a once-a-day routine — the 1-hour window is gone long before the next fire — but it changes the math when the cadence and the cache TTL line up.

If a Routine doesn't fit

Two adjacent options sit beside Routines, and either may be the right answer:

• Stay in an interactive session. Hands-on work where the cache stays warm is the cheapest mode by a wide margin. Don't reach for automation just because automation exists.
• Desktop scheduled tasks. A separate feature that runs scheduled work on your local machine instead of in the cloud. Useful when the task needs local files or a setup you don't want to recreate in a cloud environment — and when you don't need the "laptop closed" property a Routine gives you.

The decision isn't local vs. cloud; it's does this work want a human in the loop or not. A Routine is the right answer when the honest answer is no.

Sources: Automate work with routines (product name, autonomy, subscription billing, schedule minimums, trigger types); Prompt caching (5-minute default TTL, 1-hour tier, 1.25×/2× write multipliers, 0.1× read multiplier); Schedule recurring tasks in Claude Code Desktop (local-scheduled alternative). Verified 2026-06-24.