So far
$793.08
into AWS, 1 October 2026 to 4 November 2026, sales tax of $73.79 included
21–67 h
of my own time at the keyboard, over 36 days from 1 October 2026
5
costs not counted yet, because only I can supply them: Claude subscription (plan and months); Direct Anthropic API spend; landmarks.city domain and Route 53 zone (management account); GitHub (Build-Fractal plan); Other tools and hardware
Money
All of the cloud work runs in one AWS account kept for this project, so its bill is the cost. The figures are AWS's own, from Cost Explorer.
By month
| Month | AWS | Bill |
|---|---|---|
| October 2026 | $781.65 | closed |
| November 2026 | $11.43 | to 4 November 2026, estimated until the bill closes |
| Total | $793.08 |
Nothing before October 2026. The bill shows usage and tax only: no credits, no refunds. The busiest day was 20 October 2026 ($168.34); the two most recent days in the bill, 3 November 2026 and 4 November 2026, cost $1.46 and $1.45.
By service
| Service | Cost | Share |
|---|---|---|
| SageMaker (GPU and CPU jobs) | $408.24 | 51% |
| S3 storage and requests | $224.63 | 28% |
| Sales tax | $73.79 | 9% |
| Bedrock: Claude Haiku 4.5 | $38.81 | 5% |
| Bedrock: Claude Sonnet 4.6 | $17.10 | 2% |
| Lambda | $10.53 | 1% |
| Bedrock: Claude Opus 4.6 | $6.06 | 1% |
| WAF | $5.41 | 1% |
| Everything else (10 services: ECS, EFS, Bedrock: other models, AWS Backup, AWS Secrets Manager, ECR, SQS, AWS Cost Explorer, VPC, CloudWatch) | $8.50 | 1% |
What is inside that total
Pieces of work I can put a number on. They are already in the AWS figures above, so they are not added again.
| Work | Cost | Basis |
|---|---|---|
| Indiana building facts read (Haiku 4.5) | $19.00 | measured |
| Form-response profiles (Haiku 4.5) | $10.13 | measured |
| Facts gold set round 1 (Sonnet adjudicator) | $2.05 | measured |
| Facts gold set round 2 | $3.18 | measured |
| Qwen3-VL PDF parser test | $0.89 | measured |
| Twelve failed page-image jobs | $0.88 | measured |
| Route tile build (one build) | $1.75 | estimated |
| Hedge re-read: Arkansas pilot (Sonnet 4.6) | $2.26 | measured |
| Hedge re-read: Arkansas under the district rule (Sonnet 4.6) | $2.23 | measured |
| Page embeddings and trials: GPU training (g6e; 126.3 spot hours $359.49 and 5.5 on-demand hours $17.23) | $376.72 | measured |
| Page images and page copy: CPU jobs (c7i; 53.5 hours) | $22.93 | measured |
| Glacier Instant Retrieval read (3939 GB) | $118.17 | measured |
Done for nothing: national listing embedding (Ollama on the Mac); national style read (Claude subscription).
Outside AWS
| What | Cost |
|---|---|
| Claude subscription (plan and months) | not yet supplied |
| Direct Anthropic API spend | not yet supplied |
| landmarks.city domain and Route 53 zone (management account) | not yet supplied |
| GitHub (Build-Fractal plan) | not yet supplied |
| Other tools and hardware | not yet supplied |
Time
36 days on the calendar, 1 October 2026 to 5 November 2026. I typed 369 messages to Claude on 26 of them. Messages close together make one sitting, and each sitting counts its length plus 5 minutes:
| A sitting ends after a pause of | Hours |
|---|---|
| 10 minutes | 20.8 |
| 20 minutes | 24.9 |
| 30 minutes | 29.6 |
| 60 minutes | 46.9 |
So 21 to 67 hours. The top of the range adds the commits on 5 early days whose message logs are thin. It is an undercount: the early logs kept only about 20 of the roughly 140 messages of the first week, messages sent from my phone are not counted, and neither is the time spent testing, reading and thinking.
The agents worked longer than I did: commits and pull requests, theirs and mine, fall on 29 days and span 36 to 84 hours by the same rule. That is how long work was under way, not anyone’s hours.
What got better, and what didn't
Each round of work, with the number it moved. The misses are here too: where a projection was wrong, or something cost more than it should have.
1 November 2026Style read
Arkansas re-read under the district rule
Before: 17 firm, 173 influence, 3 not. After: 24 firm, 165 influence, 3 not, 1 left for a person.
Gold check: firm 74.5% to 75.0%. About 23% of the remaining claims still read as influence, so the sweep is smaller than the problem. 8 of the 165 influence claims have a form field the exact match missed. The Sanders House was wrongly sent to influence and was corrected by hand. $2.23.
1 November 2026Miss
A duplicate copy of the PDFs, now deleted
Before: 3,450 GB in Standard storage; the account about $3.75 a day (27-31 October). After: 397 GB; about $1.45 a day (3-4 November), about $44 a month.
88,154 PDFs staged for the page run stayed in Standard storage until 1 November, about $2.10 a day while they stood.
31 October 2026Miss
The page-embedding run cost three to four times its projection
Before: projected $90-115 on spot for about 110 GPU-hours. After: $376.72 for 131.8 GPU-hours.
The hours were close; the price was not. Managed spot on ml.g6e.xlarge cost $2.61 an hour against $2.62 on demand. The run finished at 3,543,706 page vectors for all 50 states and DC.
20 October 2026Miss
A Glacier read cost $118
Result: 3,939 GB retrieved from Glacier Instant Retrieval, $118.17.
Mostly on 20 October, as the PDFs were copied out of the archive for the page run.
16 October 2026Search
The search planner gets every test sentence right
Before: 53 of 64 exact; precision 0.992, recall 0.910. After: 64 of 64 exact; precision 1.000, recall 1.000.
16 October 2026Style read
Arkansas hedged styles re-read (pilot)
Result: of 193 claims re-read: 17 firm, 173 influence, 3 not.
Sonnet 4.6 on Bedrock, $2.26.
15 October 2026Style read
Every state and DC read for style
Before: 25 states. After: all 50 states and DC.
Run on the Claude subscription, not billed to AWS.
15 October 2026Building facts
Indiana facts, second round
Result: style precision 0.82, significant dates 0.96.
Built years, significant persons and alterations are still below their bars.
10 October 2026Style read
Half the country's nominations read for style
Before: the National Register's own style field. After: 25 states read, 38,969 nomination texts.
7 October 2026Search
Listings embedded on the Mac instead of an endpoint
Before: a SageMaker endpoint, projected at $57-60 nationally (Wyoming pilot, 499 listings). After: the same model in Ollama on the Mac, $0; vectors matched the endpoint's at mean cosine 0.998.
The story coming soon
How the index came to be, week by week, from the first question about Columbus, Indiana to a search over every page of every nomination.
Corrections
If a number here looks wrong, say so and it will be checked against the bill. Credits names everyone whose work the index rests on, and the roadmap says what comes next.