Coding agents cost $4,125 a month because 73% of it is context you already sent

Take the fleet apart per call. $4,125 across 30,000 requests is 13.75 cents a request ^[1], and ten of those cents are spent before the model writes a character: 20,000 input tokens at $5 per million ^[2][2]. The generated output is 3.75 cents of it ^[2]. That is what the 73% figure means in practice ^[6], and it is why approval gates and per-team limits arrive after the money is gone.

Anthropic bills cache reads at a tenth of base input, with writes at 1.25x for a five-minute window or 2x for an hour ^[7]. The post's snippet declares 40% of input a stable prefix. Run the terms as written: 240 MTok at 0.1x is $120, the remaining 360 MTok at full rate is $1,800, output stays at $1,125, total $3,045 ^[3]. That is a 26% cut ^[4], not the roughly 9% the prose claims ^[8]. The number printed under the code, $3,765, is exactly what a 9% saving looks like ^[5], so the comments and the conclusion cannot both be right. The snippet also prices no cache writes at all ^[10], and writes are billed above base rate ^[7], so whatever the correct figure is, prefix churn pulls it down.

The routing number needs no reconciliation. Same fleet, same context, Haiku 4.5: $825 ^[10], which is 80% off and $3,300 a month back ^[6]. Divide the Opus bill by five and each agent costs $825 ^[7], so one Opus agent costs what the whole fleet costs on Haiku. Add the 50% batch discount that both major providers offer ^[11] on latency-tolerant traffic and the same work lands at $412.50 ^[8].

Price levers stop paying eventually, because retrieval payloads grow with document count: more documents match ^[12]. The post's example is four chunks totalling about 7,000 tokens, two of them contradicting each other, against 200 tokens of resolved facts carrying validity dates and provenance ^[12]. The contradiction is the part that costs money, since it is what makes an agent retry. On the vendor's own Terminal-Bench 2.1 run, a task-scoped memory layer used 41.2% fewer tokens at 72.6% lower cost, with mean reward moving from 83.37% to 88.31% across 445 trials ^[13], a gain of 4.94 points or 5.9% relative ^[9]. Cost fell further than tokens because fewer retries and shorter runs compound with smaller payloads ^[14]. The author says to read the methodology rather than trust the number ^[13], which is the right instruction to take literally.

The one step whose payoff does not depend on which of these figures is correct is the bucket split: system prompt, retrieved context, conversation history, output ^[15]. A day of work, and it tells you which of the four you have been arguing about for no reason.