Why session length is the cost
Every message in a Claude Code session re-reads the whole conversation so far. So the hundredth turn of a session costs far more than the tenth, even when the answer is the same size.
On the 1M model, a session can keep growing well past 200,000 tokens before anything compacts it, and every turn pays again for everything above it.
What August looked like
In late August I audited my own usage from the session transcripts. 71.4% of all the context Claude read happened on turns where the conversation was already past 200,000 tokens. The largest single session peaked at 777,761 tokens.
Those two figures are as I recorded them at the time. Claude Code deletes transcripts after 30 days by default, so the data behind them is gone and they cannot be re-run today.
The change: one setting
Claude Code compacts a session when its context reaches a set size. It summarises the conversation and carries on from the summary. You can set that size yourself with one command:
/autocompact 200kIt saves this line to your settings file, and every session from then on uses it:
"autoCompactWindow": 200000I set it to 300,000 on 28 August, then lowered it to 200,000 on 10 September. The command takes anything from 100k to 1M, and auto puts it back to the default.
What a month under the cap shows
These numbers cover 14 September to 15 October 2026, with the cap at 200,000. That is 1,229 sessions and 74,398 turns. A session here is a conversation that I or a scheduled agent started. The helper agents a session spawns run in their own context, so they count towards the turns but not the sessions.
Compaction fires before the window is full. With the window at 200,000, the median compaction happened at about 168,000 tokens. On every day but one, no session went past 185k.
Counted the way the August audit counted, my own sessions in VS Code only, turns past 200,000 tokens now account for 0.006% of the context Claude reads. In August, as recorded at the time, that figure was 71.4%. With scheduled agents and helper agents included it is 0.022%.
What it costs you
Each compaction is a summary, so some detail from early in the session is lost. On a long task you will sometimes need to restate a decision the summary dropped.
Compacting is also a large request in its own right, because it reads the whole conversation to summarise it. When a task is finished rather than just long, starting a fresh session with /clear is cheaper than letting it compact.
What I can't tell you
I don't have a clean before-and-after on cost. The transcripts from before the cap were deleted by the 30-day cleanup, and the models I use changed over the same weeks. Any cost comparison would mix the cap up with those other changes.
If you plan to measure a change like this on your own setup, raise cleanupPeriodDays in your settings first, so the before is still there when you want to compare.
Claude