AI AutomationFounder NotesProductivity

Two clocks on your Claude account: the 5-hour window and the weekly cap

·5 min read

I spent the first month of using Claude treating usage limits like weather — something that just happened to me at inconvenient times. Then I opened Settings → Usage one afternoon, looked at the two progress bars, and stopped being surprised by them.

There are two clocks running on your account at once. Once you can see them, the math gets simple.

What's actually being measured

The first one is the five-hour session window. It's not midnight-to-5am; it's a rolling five-hour block that starts the moment you send your first message. As long as you're inside that window, every message adds up against a per-session cap. Once five hours have passed since that first message, the session resets and the counter goes back to zero.

The second one is the weekly cap. Anthropic introduced this in late August 2025, and on the current Max plan it's actually two separate bars — one for Opus-only usage, and one for everything else (Sonnet, Haiku, and the smaller models). Either of them hitting 100% throttles you, regardless of how quiet your 5-hour window looks.

Most of the surprise I see in other people's setups comes from assuming these are the same meter. They're not. They're two sums running in parallel, and either one can be the thing that hits you first.

If you want a concrete number to anchor on: on May 6, 2026, Anthropic doubled Claude Code's 5-hour caps for Pro/Max/Team/Enterprise and dropped the peak-hours reduction that used to make weekday mornings the worst time to use Opus. If you're comparing today's post to one you read three months ago, the comparison isn't fair.

What I don't trust is any specific "X messages per 5 hours" number anyone quotes — including my own. Anthropic adjusts these when compute gets cheaper, when a new model lands, or when a partnership like the SpaceX one comes online. The shape of the two-clock system is the durable thing; the exact numbers aren't.

Two minimalist analog clock faces drawn side by side on a warm cream background — one slightly larger and lifted, one smaller and lower, both with thin tick marks and no numerals, each with a thin red second hand
Two minimalist analog clock faces drawn side by side on a warm cream background — one slightly larger and lifted, one smaller and lower, both with thin tick marks and no numerals, each with a thin red second hand

What counts (and what doesn't)

A few things that move the meter, in roughly descending order of cost:

  • Message length. A 2,000-token prompt counts more than a 50-token one.
  • File attachments. Larger files count more; re-uploading the same file in a fresh chat counts it again.
  • Model choice. Opus > Sonnet > Haiku, with effort level (the way you set reasoning depth in Claude Code) sitting on top of that.
  • Tool usage. Web search, Research, and artifact-heavy work all count.
  • Conversation length. Long chats get more expensive as context accumulates.

A single minimalist clock face on a warm cream background with a deep-red pie-slice wedge inside the dial marking the "elapsed portion" of the rolling 5-hour session window — only tick marks, no numerals, no text
A single minimalist clock face on a warm cream background with a deep-red pie-slice wedge inside the dial marking the "elapsed portion" of the rolling 5-hour session window — only tick marks, no numerals, no text

And a few things that don't — or that count far less than people expect:

  • Project caching. Drop a doc into a Project and it gets cached; reusing it costs almost nothing. This is by far the biggest lever most people under-use.

A single flat vector folder icon on a warm cream background — a near-black rounded rectangle with one small folded corner tab at the top-right and a small deep-red square accent in the corner, representing cached or reusable saved content
A single flat vector folder icon on a warm cream background — a near-black rounded rectangle with one small folded corner tab at the top-right and a small deep-red square accent in the corner, representing cached or reusable saved content

  • In-chat context. Claude remembers earlier turns within the same conversation. Referring back doesn't re-charge.
  • Re-asking similar prompts. Frequently-used prompts get partial caching.

The Projects trick alone probably saved me a third of my weekly Opus budget once I started using it. If you've been uploading the same brand brief or product doc into a fresh chat every time, that's probably been the single biggest leak to fix.

How I plan around it

I stopped trying to "save up" prompts and started treating the two clocks as a calculator problem. Inputs: today's workload, what model the task actually needs, where I am in the weekly cap, and how much time is left in the current 5-hour session. Output: which model I open for the next task.

A few rules I actually run by:

  1. Sonnet first, Opus on demand. Default to Sonnet 4.5 for anything that isn't an architectural call. Opus only when I genuinely need it.
  2. Project everything that repeats. Specs, codebases, client briefs, my own templates — anything I'd re-paste tomorrow.
  3. Shorten the prompt, not the conversation. Better to start a fresh chat than to drag a 40-turn thread into something Opus has to reread.
  4. Check the bars at 10am and 4pm. Two cheap glances keep me from being surprised on a Friday afternoon.

None of this is exotic. It's the same budgeting you'd do with electricity in an off-grid cabin — two meters, day-to-day sum, week-to-week sum, and a quiet preference for not getting caught off-guard.

If you're new to Claude and the limits feel arbitrary, the fix is in two places: Settings → Usage (so you can actually see both bars), and Anthropic's usage-limit best practices page (so you can read what they themselves say counts). The mystery is mostly that nobody looks at either one.

Cover photo: an analog clock by Larry and Teddy Page, via Wikimedia Commons (CC-BY 2.0).