What one Bureau run costs

This page answers one question: what does one Bureau run cost, in tokens and API-equivalent dollars, by role, across the runs that produced devweb's essays, glossary entries and topic overviews? Across the two essay runs where every leg was captured, the totals are 24.64 and 12.27 in API-equivalent US dollars, and in both of them the Conductor is the single largest line: 18.81 of the 24.64, and 6.51 of the 12.27. Every figure below is copied from the run accounting records as they were written, including the rows that recorded nothing.

Price basis

Alias	Priced as	Input $/M	Cache write $/M	Cache read $/M	Output $/M	Source	Fetched
sonnet	Claude Sonnet 5	2.0	4.0	0.2	10.0	https://platform.claude.com/docs/en/about-claude/pricing	2026-09-06
opus	Claude Opus 5	5.0	10.0	0.5	25.0	https://platform.claude.com/docs/en/about-claude/pricing	2026-09-06
haiku	Claude Haiku 4.5	1.0	2.0	0.1	5.0	https://platform.claude.com/docs/en/about-claude/pricing	2026-09-06
fable	Claude Fable 5.1	10.0	20.0	0.25	50.0	https://platform.claude.com/docs/en/about-claude/pricing	2026-09-06
grok (OpenRouter)	x-ai/grok-4.3	1.25	n/a	0.2	2.5	https://openrouter.ai/x-ai/grok-4.3	2026-09-06

"API-equivalent" is carrying weight on this page, so here is what it means.

The Claude legs ran on a Claude Code subscription. They were not metered per token, and there is no invoice for them. Their dollar figures are modeled: the token counts each run recorded, multiplied by the list prices in the table above. The only legs that were actually metered are the Grok cross-model passes, which ran through OpenRouter and have a real per-call price.

Four more details that change how to read the numbers:

• Grok token counts are estimated from byte counts at 4 bytes per token. They are marked (est) in the leg tables below, and they are not measurements.
• Cache writes are priced at the 1-hour cache rate.
• The model aliases (sonnet, opus, haiku, fable) are priced at the model each alias resolves to on the as-of date, 2026-09-06. Runs from June and July executed on whatever those aliases pointed at then, which may have been earlier models at different prices.
• The token counts include cache reads, which is why a single essay run reports 27,804,254 tokens processed.

What these dollars are not

Three claims this page does not make.

Not spend. Nobody paid 24.64 for the run that produced /articles/vectors-in-the-file. That is what the recorded tokens would have cost at API list price on the table above. A subscription was the actual billing relationship, and a modeled equivalent is not an invoice.

Not comparable across providers or generations. Claude legs are priced on the Claude list, Grok legs on the OpenRouter list, and both are read as of one day. Putting a June run and a September run in the same column does not put them on the same price curve.

Not an average. Sixteen runs are in the table. Three are graded complete coverage, and two of those are essay runs. Two data points do not support an average, a per-article cost, or a trend.

The runs

Run	Date	Workflow	Produced	Coverage	Legs captured/recorded	Specialists recorded	Tokens processed (captured)	Output tokens	Cache-read share	Grok calls	Grok USD	Claude USD equiv	Total USD equiv
20260723-devweb-recallatron-improvements	2026-07-23	write-article	structured-memory-graph-workers	partial	10/11	10	3,833,106	15,531	88%	1	0.0145	4.92	4.93
20260723-devweb-static-nextjs-wordpress	2026-07-23	write-article	static-nextjs-vs-wordpress	none	0/1	0	0	0	n/a	1	0.0088	n/a	n/a
20260905-glossary-harness-cluster	2026-09-05	write-glossary	8 glossary-terms	partial	3/3	0	35,429,140	63,479	94%	8	0.0184	21.77	21.79
20260906-glossary-memory-cluster	2026-09-06	write-glossary	7 glossary-terms	partial	3/3	0	55,082,076	42,399	95%	7	0.0166	26.03	26.05
20260906-topic-overviews	2026-09-06	write-topic-overview	3 topic-overviews	complete	11/11	8	50,805,095	33,626	93%	3	0.0096	34.25	34.26
20260624-delegate-gating-article	2026-06-24	write-article	gate-the-flow-not-the-judgment	none	0/9	9	0	0	n/a	1	0.0105	n/a	n/a
20260624-meta-pipeline-article	2026-06-24	write-article	the-pipeline-that-wrote-this	none	0/9	9	0	0	n/a	1	0.0113	n/a	n/a
20260624-project-agent-article	2026-06-24	write-article	deny-by-default-grounding	none	0/9	9	0	0	n/a	1	0.0100	n/a	n/a
20260625-growoperative-architecture-article	2026-06-25	write-article	growoperative-foaf-protocol-architecture	none	0/9	9	0	0	n/a	1	0.0150	n/a	n/a
20260626-rheo-memory-v2-article	2026-06-26	write-article	give-the-model-the-tools-not-the-context	none	0/9	9	0	0	n/a	1	0.0098	n/a	n/a
20260705-article-rheo-track5	2026-07-05	write-article	vectors-in-the-file	complete	12/12	11	27,804,254	105,283	95%	1	0.0106	24.63	24.64
20260705-article-run-metrics	2026-07-05	write-article	agent-run-cost-attribution	partial	11/12	11	4,466,012	16,751	82%	1	0.0116	7.12	7.13
20260712-audit-convergence-article	2026-07-12	write-article	clean-audit-run-proves-nothing	complete	12/12	11	12,955,270	45,980	95%	1	0.0111	12.26	12.27
20260723-bureau-shift-left-quality	2026-07-23	write-article	shift-left-quality-agentic-pipeline	none	0/0	0	0	0	n/a	1	0.0092	n/a	n/a
20260723-devweb-wordpress-nextjs-v2	2026-07-23	write-article	static-nextjs-vs-wordpress	partial	1/1	0	112,104	511	99%	1	0.0109	0.08	0.09
20260723-write-article-evalpal	2026-07-23	write-article	evalpal-audhd-assessment	partial	2/2	0	8,539,517	9,323	95%	1	0.0097	8.26	8.27

Coverage is a grade on the record, not on the run. A run grades complete when the Conductor and at least one specialist spawn are recorded, and every recorded leg carries exact token counts. Partial means only some of the recorded legs carry counts, so the totals in that row are a floor and not the run's real footprint. None means no leg carries counts at all, which is why those rows show 0 tokens and n/a dollars: there is nothing to model, and filling the gap with an estimate would make the table look better than the record is.

Two columns need reading carefully. Legs captured/recorded counts only the legs the accounting actually recorded, so a leg the run never recorded stays out of the denominator and shows as not recorded in the per-leg table below. The Delegate and the cold reviewer on runs that predate the Delegate are the common case. Specialists recorded is the number of spawn records present in the accounting, not the number of specialists that ran: a run whose accounting recorded no specialist spawns still ran them, so its totals undercount by whatever those legs spent.

The per-leg detail behind the grade:

Run	Schema	Coverage	Conductor	Delegate	Reviewer	Specialists captured / recorded	Grok calls
20260723-devweb-recallatron-improvements	v2	partial	not recorded	exact	not recorded	9/10	1
20260723-devweb-static-nextjs-wordpress	v2	none	not recorded	partial	not recorded	0/0	1
20260905-glossary-harness-cluster	v2	partial	exact	exact	exact	0/0	8
20260906-glossary-memory-cluster	v2	partial	exact	exact	exact	0/0	7
20260906-topic-overviews	v2	complete	exact	exact	exact	8/8	3
20260624-delegate-gating-article	v1	none	absent	absent	absent	0/9	1
20260624-meta-pipeline-article	v1	none	absent	absent	absent	0/9	1
20260624-project-agent-article	v1	none	absent	absent	absent	0/9	1
20260625-growoperative-architecture-article	v1	none	absent	absent	absent	0/9	1
20260626-rheo-memory-v2-article	v1	none	absent	absent	absent	0/9	1
20260705-article-rheo-track5	v2	complete	exact	absent	absent	11/11	1
20260705-article-run-metrics	v2	partial	partial	absent	absent	11/11	1
20260712-audit-convergence-article	v2	complete	exact	not recorded	not recorded	11/11	1
20260723-bureau-shift-left-quality	v1	none	absent	absent	absent	0/0	1
20260723-devweb-wordpress-nextjs-v2	v2	partial	not recorded	exact	not recorded	0/0	1
20260723-write-article-evalpal	v2	partial	exact	exact	not recorded	0/0	1

The Schema column explains most of the empty rows. Six runs are schema v1, which recorded no per-leg token data at all: the five from June 2026, plus 20260723-bureau-shift-left-quality. Those articles exist and shipped, but the runs that made them left no accounting behind. One more row, 20260723-devweb-static-nextjs-wordpress, is schema v2 and still captured nothing usable, at 0 of the 1 leg it recorded. The remaining six v2 rows record something without clearing the bar, and they fall short in two different ways. 20260723-devweb-recallatron-improvements and 20260705-article-run-metrics each miss one of the legs they recorded, at 10 of 11 and 11 of 12. The other four captured every leg they recorded and still grade partial: none of them recorded a specialist spawn at all, which is what the 0 under Specialists recorded means. Those runs spawned specialists; the record simply does not hold them, so their totals are a floor.

Three rows are graded complete: the topic-overview cluster, which produced three topic overviews rather than one essay, and the two essay runs broken out below.

Where the cost sits

20260705-article-rheo-track5, which produced /articles/vectors-in-the-file (2026-07-05)

Agent	Model	Input	Cache write	Cache read	Output	USD equiv	Arithmetic
The Conductor	opus	26,512	470,069	23,346,127	92,070	18.81	0.027M x $5.0 + 0.470M x $10.0 + 23.346M x $0.5 + 0.092M x $25.0 = $18.81
Tally	sonnet	20	111,894	1,305,545	10,460	0.81	0.000M x $2.0 + 0.112M x $4.0 + 1.306M x $0.2 + 0.010M x $10.0 = $0.81
The Counselor	sonnet	10	43,730	143,002	317	0.21	0.000M x $2.0 + 0.044M x $4.0 + 0.143M x $0.2 + 0.000M x $10.0 = $0.21
The Scribe	sonnet	9	43,169	229,164	368	0.22	0.000M x $2.0 + 0.043M x $4.0 + 0.229M x $0.2 + 0.000M x $10.0 = $0.22
The Scribe	opus	13,116	73,339	283,394	33	0.94	0.013M x $5.0 + 0.073M x $10.0 + 0.283M x $0.5 + 0.000M x $25.0 = $0.94
The Scribe	opus	13,114	71,116	221,730	18	0.89	0.013M x $5.0 + 0.071M x $10.0 + 0.222M x $0.5 + 0.000M x $25.0 = $0.89
The Scribe	opus	13,116	64,295	273,956	28	0.85	0.013M x $5.0 + 0.064M x $10.0 + 0.274M x $0.5 + 0.000M x $25.0 = $0.85
The Scribe	opus	13,114	46,466	164,261	20	0.61	0.013M x $5.0 + 0.046M x $10.0 + 0.164M x $0.5 + 0.000M x $25.0 = $0.61
The Counselor	sonnet	13	49,607	324,452	767	0.27	0.000M x $2.0 + 0.050M x $4.0 + 0.324M x $0.2 + 0.001M x $10.0 = $0.27
The Counselor	sonnet	8	89,606	71,270	723	0.38	0.000M x $2.0 + 0.090M x $4.0 + 0.071M x $0.2 + 0.001M x $10.0 = $0.38
The Scribe	sonnet	5	28,189	61,909	17	0.13	0.000M x $2.0 + 0.028M x $4.0 + 0.062M x $0.2 + 0.000M x $10.0 = $0.13
The Challenger	opus	13,110	41,051	49,483	462	0.51	0.013M x $5.0 + 0.041M x $10.0 + 0.049M x $0.5 + 0.000M x $25.0 = $0.51
Grok pass 1	grok-4.3	~2,829 (est)	n/a	n/a	~2,822 (est)	0.0106	(11315/4)/1e6 x $1.25 + (11287/4)/1e6 x $2.5 = $0.0106
Total						24.64	Claude 24.63 + Grok 0.0106

20260712-audit-convergence-article, which produced /articles/clean-audit-run-proves-nothing (2026-07-12)

Agent	Model	Input	Cache write	Cache read	Output	USD equiv	Arithmetic
The Conductor	opus	874	71,648	9,701,853	37,519	6.51	0.001M x $5.0 + 0.072M x $10.0 + 9.702M x $0.5 + 0.038M x $25.0 = $6.51
Tally	sonnet	6	56,367	116,868	383	0.25	0.000M x $2.0 + 0.056M x $4.0 + 0.117M x $0.2 + 0.000M x $10.0 = $0.25
The Counselor	sonnet	10	33,041	161,438	571	0.17	0.000M x $2.0 + 0.033M x $4.0 + 0.161M x $0.2 + 0.001M x $10.0 = $0.17
The Scribe	sonnet	7	29,811	129,168	365	0.15	0.000M x $2.0 + 0.030M x $4.0 + 0.129M x $0.2 + 0.000M x $10.0 = $0.15
The Scribe	fable	13,353	72,425	202,474	1,171	1.69	0.013M x $10.0 + 0.072M x $20.0 + 0.202M x $0.25 + 0.001M x $50.0 = $1.69
The Scribe	opus	13,371	57,781	655,368	1,899	1.02	0.013M x $5.0 + 0.058M x $10.0 + 0.655M x $0.5 + 0.002M x $25.0 = $1.02
The Scribe	opus	13,353	46,824	160,734	659	0.63	0.013M x $5.0 + 0.047M x $10.0 + 0.161M x $0.5 + 0.001M x $25.0 = $0.63
The Scribe	opus	13,357	48,296	268,958	888	0.71	0.013M x $5.0 + 0.048M x $10.0 + 0.269M x $0.5 + 0.001M x $25.0 = $0.71
The Counselor	sonnet	15	46,761	414,010	108	0.27	0.000M x $2.0 + 0.047M x $4.0 + 0.414M x $0.2 + 0.000M x $10.0 = $0.27
The Counselor	sonnet	9	31,777	162,471	655	0.17	0.000M x $2.0 + 0.032M x $4.0 + 0.162M x $0.2 + 0.001M x $10.0 = $0.17
The Scribe	sonnet	10	31,433	252,858	800	0.18	0.000M x $2.0 + 0.031M x $4.0 + 0.253M x $0.2 + 0.001M x $10.0 = $0.18
The Challenger	opus	13,349	39,427	49,785	962	0.51	0.013M x $5.0 + 0.039M x $10.0 + 0.050M x $0.5 + 0.001M x $25.0 = $0.51
Grok pass 1	grok-4.3	~2,958 (est)	n/a	n/a	~2,958 (est)	0.0111	(11833/4)/1e6 x $1.25 + (11832/4)/1e6 x $2.5 = $0.0111
Total						12.27	Claude 12.26 + Grok 0.0111

Two things stand out, and they are the same thing seen twice.

The orchestrating leg dominates. In the first run the Conductor models out at 18.81 against a run total of 24.64; the next largest single leg is a Scribe opus pass at 0.94. In the second run the Conductor is 6.51 against a total of 12.27, and the next largest leg is the Scribe's fable pass at 1.69. Across both runs the dataset attributes 25.32 to the Conductor, 8.02 to the Scribe, 1.47 to the Counselor, 1.06 to Tally and 1.02 to the Challenger. The specialists are cheap. The agent holding the whole run in context is not.

Cache reads dominate the tokens. The Conductor's line in the first run is 23,346,127 cache-read tokens against 26,512 input and 92,070 output; in the second it is 9,701,853 against 874 and 37,519. The Cache-read share column reports 95% for both runs, and the dataset's rollup gives 0.951 across the pair. That is also why the dollar figures stay as low as they do: on the price table, an opus cache read is 0.5 per million against 5.0 for input and 25.0 for output. A run can process 27,804,254 tokens and still model out at 24.64.

What this cannot say yet

The rollup, for the two complete-coverage essay runs only:

• complete_essay_runs: 2
• median_usd_equiv_per_essay: 18.45
• min_usd_equiv: 12.27
• max_usd_equiv: 24.64
• median_tokens_processed: 20379762
• cache_read_share: 0.951

A median across two runs is the midpoint between them. 18.45 is not a typical run cost, because two observations cannot establish what typical is, and the two observations here are 12.27 and 24.64 on the same workflow. Read the rollup as a range that has been seen, not as a forecast.

The rest of the table is thinner still. Nine of the sixteen rows carry any token data at all, and six of those nine are graded partial, which means their totals undercount by an unknown amount. The other seven rows have nothing to report. Nothing here supports a per-article price, a cost-per-word, a comparison between workflows, or a claim that runs are getting cheaper or more expensive over time. The one comparison the data does support is within a run: which role spends, and on what.

Coverage improves as more runs complete under the current accounting schema. This page is generated from the run records and regenerated as more of them land, so the sample it reports on is whatever had been captured on the as-of date, 2026-09-06. Related reading: agent run cost attribution on how the per-leg numbers get collected, and multi-agent pipeline for what the roles in the leg tables actually do.