Skip to main content

Keep going, compact, or switch models?

Every turn re-reads your whole conversation, so a big context gets expensive. Compacting shrinks it and switching to a cheaper model lowers the price of every turn, but each costs something once. Here is when that pays off.

The short version

  • Under about 50K tokens, don't compact. It takes around thirty turns to pay for itself.
  • Past about 200K, with more than a dozen turns of work left, compact. Every turn re-reads everything, so a big context is the cost.
  • Switch to a cheaper model early, not late. Switching re-writes your whole context to the new model's cache, so the bigger the context, the longer it takes to earn that back. Compacting first shrinks the bill.
  • If you've been away longer than the cache lasts, act now. The next turn pays to rebuild the whole context anyway, so compacting or switching first costs almost nothing extra.
  • Starting a different task? Just clear. There's nothing to compare, because you don't need the old context at all.

How long each move takes to pay for itself on Claude Opus 5.5, when every turn adds 5K tokens of input and 2K of output, and a 5% summary replaces the history. These are the calculator's defaults.

Turns until compacting, or switching to Claude Sonnet 5.5, is cheaper than keeping going, by context size.
Context nowCompact, cache warmCompact, cache expiredSwitch to Claude Sonnet 5.5
50K31 turnsRight away5 turns
100K18 turnsRight away9 turns
200K13 turnsRight away17 turns
400K11 turnsRight away34 turns
800K10 turnsRight away67 turns

The switch column assumes a warm cache. With an expired one, switching pays off right away too.

Check it against your session

Change the scenario to match what you're doing.

Tokens in the prefix every turn re-reads.

How much work is left. The slider stops at 120, the box at 500.

Cache right now

Cold means more than the cache lifetime has passed since your last turn.

Dollars per million input and output tokens.

The model you'd move the rest of the session to.

Assumptions Cache lifetime, summary size, and per-turn growth. The defaults are reasonable.
Cache lifetime

How long a cache write lasts. A longer one costs more to write.

A percentage of your context, or a token count over 100.

New tokens each turn adds, such as tool results and your messages.

Tokens the model writes each turn.

System prompt, tools, and project context. Compacting keeps it.

What each move costs from here

Over the next 30 turns, compact now: it comes to $4.46, against $5.89 to keep going and $6.05 to switch to Claude Sonnet 5.5.

Compacting costs $0.76 up front and pays for itself after 11 turns. Switching to Claude Sonnet 5.5 costs $1.60 up front and pays for itself after 34 turns, more than the 30 you have left.

  • Keep going
  • Compact now
  • Switch model (dashed)

Hover or use the arrow keys to read a turn. A dot marks the turn each move first costs no more than keeping going.

Start from your own session

Optional. Drop a Claude Code transcript and the scenario fills in with your real context size, turn sizes, model, cache state, and how much your compactions actually shrink the context.

Loading the session reader…

How this is calculated

Loading the explanation…