Resume this conversation? in Claude Code: when an old session eats your 5-hour limit

TL;DR The percentage in the dialog is what re-reading the saved conversation, uncached, costs against your 5-hour window, before you type a word. Pick Start a new conversation unless your next message needs the thread verbatim: the new one carries a pointer to the old transcript, and Claude reads from it when you mention earlier work. The old conversation stays in /resume.

Since v2.1.304, claude --resume on a conversation that has gone cold can stop you with "Resume this conversation?" and a percentage of your 5-hour usage limit. The reason is the one behind every expensive first turn: resuming re-sends the whole history, and after the cache lifetime (one hour on a subscription, five minutes once you are on credits) none of it is cached, so your first message processes every token as fresh input. Until now you found out from the spinner. Now Claude Code works it out before the request leaves.

What the dialog looks like

Composed from the strings in the v2.1.309 binary, with example figures; it could not be captured live, and the next section says why:

Resume this conversation?

This conversation has been inactive for 3 days and is 412,300 tokens long.
Resuming it will use about 14% of your 5-hour usage limit.

❯ Resume
  Start a new conversation
  Claude can look this one up if you refer to it later

enter to confirm · esc to resume

Three things the wording hides:

• The estimate is for the saved conversation alone. It leaves out what you have already spent in the window, so 14% on top of a window at 90% is a cutoff.
• Esc is not a cancel, it resumes.
• When the estimate reaches 100%, the line says it will use "all" of your 5-hour usage limit.

Where the number comes from

Claude Code works it out from the saved transcript:

last turn's tokens × the model's rate ÷ your plan's 5-hour budget

• The tokens are input, cache reads and cache writes together, capped at the model's context window.
• Rates and budgets arrive in a remote configuration, so Anthropic can change them without a release. The copy from October 23 lists budgets for Max 5x and Max 20x, and a rate for every current model. Pro's budget wasn't in that copy, though the changelog names Pro first.
• The dialog only shows from a floor that is 5% today.

With those figures, the model weighs more than the size. The same 400K-token conversation on Max 5x:

On Max 5x, Opus 5.5 would need 1.36 million tokens to reach 5%, more than its window holds; on Max 20x not even Fable 5.1 at a million tokens gets there. On Max, with that configuration, this is a Fable 5.1 dialog from 300K tokens up.

Two more conditions:

• The cache has to be cold, or the model has to differ from the one the conversation ended on: resuming a Sonnet conversation on Opus is a cold read even if you left two minutes ago.
• It stays quiet when you are already past the limit or drawing on usage credits, because there the price is no longer a share of a window.

What "Start a new conversation" actually does

It runs /clear, so the old conversation stays where /resume finds it, and it adds a note Claude sees and you don't:

• the old session id;
• the path of its transcript, a JSON lines file under ~/.claude/projects/;
• the last prompt you sent and the reply to it, each trimmed to 1,500 characters;
• the instruction to search or read that file when you refer to earlier work, reading only what it needs, instead of asking you to repeat it.

That's the "Claude can look this one up" line. It works because every conversation is already on disk in plain text: the lookup is a grep, not a resume.

The line goes missing when your organization's data policy forbids linking a new conversation to the one it replaces, or when the server configuration turns the pointer off. Then the new conversation starts blank, and the choice is a plain /clear.

How to choose

1. Read the percentage next to what is left

Run /usage first if you don't know where the window stands. The dialog's figure adds to that, and the 5-hour window is a rolling one: 14% spent at 9:00 stays spent until 14:00.

2. Decide by the next message, not by the old conversation

• Resume when the next message needs the thread verbatim: you are mid-review of a diff Claude produced, or the fix depends on a dozen tool results you don't want re-run.
• Start a new conversation when the task is related but new: open with a message that names the earlier work ("the auth migration from Tuesday, the two tests that were still failing") and let Claude read the transcript.

You can still come back with /resume when you need the whole thing.

3. Make the dialog rare

It only appears when the saved history is big and cold. Compact before you step away and what goes cold is a summary; above 200K tokens, idle compaction does it for you after 54 quiet minutes. And a conversation you won't return to is a /clear at the end, not a candidate for this dialog three days later.

What it does not warn about

Not seeing the dialog is not a clean bill. Under the floor the resume costs the same percentage, silently. A cold conversation on Opus 5.5 still re-reads every token at full price, which is the slow and expensive first turn you already know, and /cost lists it as a miss. The dialog is also wired to fire in a session you left open, when you type again after the cache has expired, keeping your unsent message; in the October 23 configuration that path was still behind a flag that was off, and the changelog only announces the resume case.

Reference

Official docs: The "Resume this conversation?" dialog

Requirements

• Claude Code v2.1.304 or later, in the terminal, on a Pro or Max plan. With an API key or on usage credits there is no 5-hour window to measure against, so there is no dialog.
• The rates, budgets and the 5% floor come from a remote configuration and can change; the figures above are that configuration's on October 23, 2026.