[guides] pacing
How to pace AI coding limits so you never run dry
“62% used” doesn’t tell you much on its own. With four hours left in the window, you’re fine. With forty minutes gone, you’ll be locked out by lunch. Pacing is comparing what you’ve used with how much of the window has passed, and it works the same for every AI coding tool.
Updated 5 min read
AI coding plans rarely cap you per month the way a phone plan does. They cap you per window. Claude Code and Codex reset a session window every five hours and a weekly one every week; Cursor resets a pool every week or every billing month, depending on the plan.
So the question is never just “how much is left?” It’s “will what’s left last until the reset?” Answering it takes two numbers: how much of the limit you’ve used, and how much of the window has passed.
Put both numbers on the same scale and compare them. If usage is behind the clock, you’re spending slower than the window allows. If it’s ahead, you’re spending faster, and at that rate you’ll run out before the reset.
- Behind the clocklasts to the reset
40% used · 60% of the window gone
- Ahead of the clockruns out early
55% used · 45% of the window gone
- Far aheadruns out at 50%
80% used · 40% of the window gone
The gap matters more than the total. 80% used is fine with 90% of the window gone. 30% used is a warning ten minutes in.
Your burn rate is what you’ve used divided by the time it took. The rest of the limit lasts as long as the remainder divided by that rate:
rate = used ÷ time elapsed
time left = (100% − used) ÷ rate
An example. Two hours into a 5-hour window you’ve used 60%. That’s 30% an hour, so the last 40% lasts 1 hour 20 minutes. You’d run out at 3h20m, 1h40m before the reset. Try your own numbers in the pace calculator.
A small change early in the window is worth more than a big one at the end. Ahead of the clock with three hours to go, try these:
- Start fresh conversations. Every turn resends the context, so a long session costs more per message. In Claude Code,
/clearbetween tasks and/compactwhen you’d rather keep going. - Use a smaller model for small jobs. Renames, tests and formatting don’t need the biggest model.
- Run fewer agents at once. Three parallel agents burn roughly three times as fast.
- Spread work across tools. If Claude Code is ahead of its clock and Codex is behind, move the next task.
- Take the break. Time away still counts on the clock, so a break lets it catch up with your usage.
Weekly limits move slower, so they catch people out. A week has about five working days, so spending 20% a day keeps you on pace. Spend 35% on Monday and you’ll either slow down by Wednesday or stop on Thursday.
Check the weekly limit when you start the day, not when you hit it. Provider by provider: Claude Code, Codex and Cursor.
Burnstate does this math for every limit you connect and keeps it in your menu bar. Each limit gets a bar with a white tick for the clock, a status (on track, burning fast or running out) and, when you’re ahead, a forecast like “runs out 15:04, 38m before reset”. A notice card warns you when a limit runs low. See Limits and pace and Notices.
[burnstate]
See every limit against the clock, in your menu bar.
Burnstate reads your Claude Code, Codex, Cursor and Vercel limits, marks how much of each window has passed, and tells you when you’d run out at your current pace.