Each API call is stateless. The model sees the text in the request and nothing else, and there is no persistent working memory between turns [8]. The continuity you feel in a long coding session is manufactured by re-sending: prior tool results and file reads, concatenated and shipped again on the next request [9].
Divide 1.99 billion cache reads [3] by 5,288 requests [1] and the average request carried about 376,000 tokens of previously seen context [15]. An average is the wrong shape for the thing, though. Turn three is cheap and turn 900 is not, which is why the author says the read-to-write ratio climbs inside any single long session as context accumulates [12].
The compactions read differently with that number in hand. All four cluster at the roughly 1M ceiling, which is the auto-compactor firing before the window overflows [14]. What the reset costs is 108 to 140 seconds each [6], so somewhere between seven and nine minutes of wall clock across the four events [16], plus whatever the summary failed to preserve [19]. Annoyance at the compactor is aimed at the cheap end of the bill.
The 32:1 figure is a claim about one developer's three sessions [1][4]. For it to hold on yours, the sessions have to run long enough to approach the same ceiling, tool output has to grow context on nearly every call, and the cache hits have to keep landing. A twenty-turn session that never pulls a large file into context will not produce anything close. The mechanism transfers; the multiple is workload-specific.
The dollar pair deserves the author's own asterisk. Check the stated 86%: 4,500 divided by 31,000 is 14.5%, so 85.5% [18], with a gap of about $26,500 between the cached and uncached prices [17]. The author notes he pays a flat subscription rather than API billing, never wrote that check, and used list rather than effective prices, so the pair is meaningful as a ratio and not as spend [11]. That is the right way to publish it.
The practical consequence is a sequencing rule. In a session you expect to run long, the cost of what you admit into context is set by the number of turns that follow it, not by the read that put it there.
