xnapz

Measured · one transcript, five weeks

Almost everything an AI coding agent bills you for is context you already sent.

We summed the usage record of every assistant turn in a five-week session and priced it at list rates, with prompt caching already on. Then we modeled three independent ways to stop paying for most of it. This page puts the answer first and the arithmetic underneath.

Derived from one continuous five-week Claude Code session on an AI site-building project — every assistant turn in a 256 MB transcript, summed and priced at Anthropic first-party rates.

The short version

97.2% of prompt tokens were re-reads — context the agent had already sent, sent again. measured
574:1 tokens read per token written, at the median turn. measured
$5,931.46 what the five weeks cost, at first-party rates with caching on. measured
−87.2% modeled reduction across three composed layers, to $758.37. modeled

What is measured and what is modeled

The distinction is the reason to trust the rest of the page, so it sits here rather than in a footnote.

Measured

  • Every token count — prompt, cache-read and output — summed per assistant turn.
  • The call count (13,632) and the window (25 June to 30 July 2026).
  • The baseline dollar figure: those counts multiplied by published list rates.
  • The per-turn medians: 511,891 tokens read, 891 written.

Modeled

  • All three reduction layers. Nothing was re-run under the modeled conditions.
  • The split between batchable and interactive traffic.
  • The share of turns a cheaper model could have taken.
  • Therefore the headline reduction itself, and the band around it.

A modeled figure is an argument, not a receipt. The sensitivity band below is where the argument is stress-tested.

Where the money goes

Three layers, composed. The percentages are not additive — each acts on the cost the previous layer left behind, which is why the third looks small and still matters.

Step Cost remaining This layer Cumulative Basis
As it ran $5,931.46 measured
Bundle the context $1,397.98 −76.4% −76.4% modeled
Match the model to the turn $1,011.16 −27.7% −83.0% modeled
Batch what can wait $758.37 −25.0% −87.2% modeled

Per call: $0.44 → $0.06.

The three layers

1 · Bundle the context — −76.4%

Stop re-sending the whole workspace on every turn. Assumes turns are grouped so a stable context block is sent once per group rather than once per turn, with the median group holding roughly a dozen turns. Cache reads are cheap per token; at 7.01B of them, cheap is still most of the bill.

2 · Match the model to the turn — −27.7%

Most turns are not hard turns. Route the mechanical majority — file reads, greps, edits applied from an already-made decision — to a cheaper tier, and keep the frontier model for the turns that actually reason. Assumes the split observed in the transcript holds under routing.

3 · Batch what can wait — −25.0%

The Batch API is 50% off with a 24-hour service window. Applied to the half of the traffic that was autonomous — overnight runs, sweeps, scheduled analysis — and to none of the interactive half. A discount on the schedule, not another compression trick.

Why three layers and not four

Bundling removes the re-sent input, and what it leaves is dominated by output tokens: 43% of the remaining cost is the words the model actually writes. No amount of context discipline compresses those. That is the floor, and it is why the third layer is a scheduling discount rather than another way to send less.

How much this moves if the assumptions are wrong

Each layer was re-run at the pessimistic and optimistic end of its assumption — fewer batchable turns, a worse model split, looser bundling, and the reverse. The result is a band of roughly 73% to 91%. The headline of 87.2% sits inside it, which is the point of showing it.

Method

13,632 assistant calls, 25 June to 30 July 2026: 7.01B prompt tokens in, 21.7M out, at a median of 511,891 read per 891 written.

The transcript is the raw session log written by the coding agent — one JSON record per turn, each assistant turn carrying a usage object with input, cache-creation, cache-read and output token counts. Those were summed, per model, across every assistant turn in the window, then multiplied by the published first-party rate for that model. Nothing was estimated and nothing was sampled: the sum is over every turn.

Cache reads are priced at a tenth of the base input rate and cache writes at one and a quarter. Both are included at those rates, which is why the baseline is what a real invoice would have said. The three layers were then modeled on top of that same token ledger by re-pricing it under each layer's stated assumption. No traffic was re-run, so none of the three is an observation.

The number this page refuses to headline

With prompt caching switched off, the same traffic would have priced at $41,016.24. That is not the baseline and does not appear above, because caching is on by default — it is a bill nobody receives. Quoting it would inflate the reduction claim from 87% to 98% on the strength of a counterfactual. It is here so you can see it was considered and set aside.

The project the transcript came from was a collaboration; the transcript analysis is mine. It is described rather than named because the brand is not mine to trade on.

Somebody else's numbers

A single transcript is a single transcript. Nate B. Jones reported 3.77B tokens in a single day of agentic coding on 29 July 2026, 96% of them reused context. Our five weeks ran at 97.2%; his one day ran at 96%. Two unrelated workloads landing within a point and a half of each other is the sign that this is a property of how coding agents work rather than a quirk of one transcript.

Why this is on a planning product's website

Because it is the same argument. A coding agent re-sends its whole world every turn because it has no structured plan to refer back to — so it reconstructs one, from scratch, on the way into every reply. Give the work a real plan with real state, and the context stops being something you pay to rebuild. That is what xnapz does with a job, and it is what these three layers do with a token bill.

The interactive version, with the figures drawn: /#/research.