Cost report

How cost tends to grow. Each request re-sends the conversation so far. As the session gets longer, that context usually grows, so per-request cost often creeps up — not just because of the new reply, but because the model is billed again for (mostly cached) history. At high cache-hit rates, cache reads dominate that bill.

If per-request cost stayed flat, total cost would rise linearly with request count (a straight line). If context keeps growing turn after turn, per-request cost rises roughly with session length, and total cost curves up quadratically (~N² in request count) — long agent runs get expensive faster than “N times the first turn.”

The blue line is one quadratic curve fitted to this branch’s cumulative cost (R^2 = 0.999; closer to 1 means the curve matches well), drawn through the observed range and extended 8 requests ahead (dashed) so you can ballpark where total spend is heading if growth continues (about $3.95 by request 87). Faint amber bars are the implied per-request cost on that projected tail (difference between successive points on the same curve) — not a separate model. Compaction, cache misses, fat outputs, or model switches bend the story — treat the fit as a guide to the shape, not a promise.