The Gate Costs More Than the Writing

Finding. Across four full runs of an autonomous research-and-publishing agent, checking the work cost more than producing it. The quality gate took 51.5% of all spend. Everything that produced the draft — research, angle selection, and the writing itself — took 42.8%.

That ratio held on every run in the sample, at 49%, 52%, 58% and 48%. It is not an artifact of one bad day.

The honest decomposition

“The gate” bundles two different activities, and the headline number is less interesting than the split. Grading is pure judgement: a separate process reads the draft against a written test and returns a verdict. Revision is generative work, but it only happens because the gate rejected something. Counting it as writing is defensible; so is counting it as the cost of having a standard.

So the defensible claims are these. Pure judging cost 23.7% — roughly a quarter of the budget spent on deciding whether the work was good enough. And having a standard at all, counting the rework it forces, cost 51.5%. Both numbers are larger than most people would guess, which is the point of measuring rather than estimating.

Per run

Two things are visible here that the totals hide.

Cost per run fell by 70% across four days, from $9.20 to $2.80, with no change in what the pipeline was asked to do. The savings came from prompt caching and from cutting redundant research calls. Anyone extrapolating an annual budget from an early run will be wrong by a factor of three.

And the last run spent $2.80 to publish nothing. The draft failed the gate across three cycles and was escalated rather than published. That is 12.7% of all spend in the sample producing zero output, which is the correct behaviour for a gate and still a real line item. Averaged over the five posts that did make it out, the sample cost $4.42 per published post.

Where it goes, by phase and model

The model concentration is the sharpest number in the set. The largest model ran 14.2% of the calls and 51.2% of the cost, and more than half of its spend was inside the gate. A cheap model ran three quarters of the calls for a quarter of the money. Tiering is doing most of the work in this pipeline, and the remaining budget question is almost entirely “which steps get the expensive model,” not “how many calls do we make.”

Method

What the system is. A Python orchestrator that sequences phases deterministically and shells out to a model for every generative or judgement step. It researches a topic and writes it into a note vault. It drafts a post from that. A separate process with no authoring context then grades the draft against a written authenticity test, and the pipeline revises on failure and publishes to WordPress only if the gate passes.

Sample. Four complete pipeline runs, 2026-08-03 to 2026-08-06. 158 API calls, 16,840,845 tokens, $22.09. Five posts published, one run publishing nothing.

What is counted. Every call is logged at the point of return with its phase label, model, token counts in four categories, and computed cost. Phase labels come from the orchestrator, not from inference after the fact, so the attribution of spend to gate versus draft is exact rather than estimated.

What is not counted, and this is the limitation that matters most. Only spend billed to an API key is priced. Work running on a subscription token is not billed per token, so it carries token counts but no dollar figure. 82.9% of tokens in this sample are priced; the other 17.1% are real work with no cost attached. If the unpriced calls are distributed differently across phases than the priced ones, every share above shifts. They appear to be spread across phases rather than concentrated, but the sample is too small to demonstrate that, so treat the percentages as accurate to a few points rather than to the decimal.

Excluded. One further date in the ledger, 2026-09-04, is a gate-only invocation with no pipeline run behind it — no research, no draft, seven calls. Including it would have inflated the gate share to 84% and the headline would have been wrong. It is excluded on the basis of the run state recording zero runs and no draft path.

What would change this

A larger sample, first. Four runs is enough to notice a pattern that repeats every time and not enough to put a confidence interval on it. Thirty runs would settle whether the 51.5% is stable or drifting.

A cheaper grader would move it most. Half the gate’s spend is one expensive model reading drafts; if a smaller model can return the same verdicts, the headline collapses. Nobody has shown that it can, and the whole design rests on the grader being at least as capable as the writer.

A better first draft would move it too, from the other direction. Revision is 27.8% and exists only because drafts fail. A pipeline that passed on the first cycle would look completely different, and the honest reading of these numbers is that this one does not pass on the first cycle often enough.

And if the unpriced 17.1% turns out to be concentrated in one phase, the shares change. Confirming that requires per-token pricing on subscription calls, which is not currently available.

Collected 2026-08-03 to 2026-09-04. First published 2026-09-04. Last revised 2026-09-04.