Three ways to save tokens in Claude Code

Every token you don't spend is more work you get done before Claude Code hits a wall. On a subscription that wall is real: usage resets on a rolling five-hour window, and a bloated session chews through it faster than you'd guess. These are the three habits I lean on to get more out of every window.

Before the fixes, it helps to know where the usage disappears. Three places, and each of the three moves targets one of them:

• Context size. Every turn re-sends the whole conversation as input. A session that's been running for an hour pays for its entire history on every single message you send after.
• Model tier. Fable 5 is the smartest model Anthropic ships, and it's priced like it. Running everything on it drains your limit far faster than most of the work needs.
• Idle windows. Your five-hour window resets whether you used it or not. Sleep through one and that capacity is just gone.

Run /compact and Claude Code rewrites everything so far into a short summary, then keeps going in the same session. Your recent messages, the file edits, CLAUDE.md, and your skills all survive. The long tail of back-and-forth gets squeezed down to the parts that still matter.

Why that saves so much: the model re-reads the entire conversation on every turn. A session sitting at 100k tokens pays for 100k tokens of input before it even gets to your new message. Compact resets that to a few thousand, and every turn after is cheaper.

It also keeps the model sharp. Anthropic documents an effect they call context rot: as the token count climbs, accuracy and recall drop off. A leaner context isn't only cheaper, it gets you better answers. So compacting at a clean seam — the end of a chunk of work, before you start the next thing — pays twice.

/compact

# or steer what survives the summary:
/compact focus on the auth refactor, drop the exploration

Auto-compact fires on its own when you're about to run out of room, but doing it deliberately beats letting it trigger mid-thought. Reach for it whenever you finish one task and stay in the same repo for the next.

Most tokens in agentic work get burned on execution: reading files, grepping, running tests, browsing. None of that needs the expensive model — judgment does. So put Fable 5 on the planning and let Sonnet 5 do the actual work.

Anthropic benchmarked exactly this split. A Fable 5 orchestrator with Sonnet 5 workers scores 96% of the all-Fable result at 46% of the price. The recipe in plain Claude Code: one Fable session writes the plan to PLAN.md, a fresh Sonnet session picks it up and executes. Switch models mid-session with /model. The orchestrator never greps; the workers never architect.

I wrote the full recipe — the three exact prompts and the review loop — in a separate guide:

The pattern and the benchmark charts come straight from Anthropic:

This is the lever most people miss. Your five-hour window resets on a rolling basis whether or not you were awake to use it. A routine — a scheduled cloud agent — runs on a cron timer without your terminal open, on Anthropic's infrastructure. Point it at the recurring jobs you'd do anyway and let them fire while you sleep, using a window that would otherwise go to waste.

You set one up from inside Claude Code with /schedule. Describe the job in plain English and when it should run, and it wires up the cron agent for you:

Every night at 2am, pull the latest commits on main, run the test suite, and write a short digest of what changed and anything that broke to nightly-digest.md. I want it waiting for me when I wake up.

Good candidates are anything on a cadence: a nightly repo digest, a dependency-and-security sweep, a daily competitor scrape, or the morning newspaper build. The work happens in hours you weren't using, and the output is sitting there at breakfast.

The three compound. Compact keeps each session lean, the model split keeps the expensive tier off the boring work, and scheduling moves that boring work into hours you weren't spending. Run one session the old way and one with all three on, and the difference in how far you get before a limit is hard to miss.

• Make /compact focus on… muscle memory at the end of every chunk of work.
• Read the orchestrator guide and save its three prompts as a skill so /orchestrate kicks off the whole flow.
• Pick one recurring job and schedule it tonight.