The weekly report

Every week the factory reports how it worked: delivery, quality, review, costs and what each dollar bought, against the previous week and industry benchmarks. Every number comes from a query against BigQuery, not from a model’s memory, so you can make business decisions on it.

As of 18 October 2026Based on 195,000 lines of code and more than 30 billion tokens burned.

What a dollar buys

Professional work per $1,000, at full list prices: no subscription discount, no cloud credits.

Hours of professional work per $1,000Person-hours, one week

Team of 5 in Poland, no AI30 h
Same team with AI assistants34 h
Rational Trains + 70-hour founder week493 h
Rational Trains, lower estimate895 h
Rational Trains, own formula1,431 h

Team: 5 people at €30 an hour. With AI: +17.5% productivity (middle of Stanford’s net +15–20%) and a $200 monthly subscription per person. Rational Trains: 2,586 to 4,134 hours of equivalent work for $2,889 at list prices.

Show as table

Hours of professional work per $1,000
Item	Hours
Team of 5 in Poland, no AI	30 h
Same team with AI assistants	34 h
Rational Trains + 70-hour founder week	493 h
Rational Trains, lower estimate	895 h
Rational Trains, own formula	1,431 h

Cost of 1,000 delivered lines of codeWithout tests, US dollars, one week

Team, no AI (QSM: 414 lines per person-month)$12,386
Team with AI assistants$10,902
Rational Trains + founder, adjusted 2.6×$746
Rational Trains, adjusted 2.6× for AI verbosity$410
Rational Trains, unadjusted$158

AI-written pull requests are 2.6 times larger (LinearB 2026), so the adjusted figures count an AI line as 1/2.6 of a human one.

Show as table

Cost of 1,000 delivered lines of code
Item	Per 1,000 lines
Team, no AI (QSM: 414 lines per person-month)	$12,386
Team with AI assistants	$10,902
Rational Trains + founder, adjusted 2.6×	$746
Rational Trains, adjusted 2.6× for AI verbosity	$410
Rational Trains, unadjusted	$158

Delivery

What left the depot this week, and how fast.

cars delivered in 6 days: 136; +36% on the previous week, in 19 releases.
median from GO to ready: 36 min; 47 minutes to merge. Up to 4 parallel lanes.
cars overnight, unattended: 21 of 21; One train, 10 hours, nobody watching.

Pull requests merged per weekSince the first day of the project

Show as table

Pull requests merged per week
Week	Merged pull requests
1	11
2	23
3	100
4	52
5	142
6	158

Cars delivered per dayMerged car pull requests, over two weeks

Show as table

Cars delivered per day
Day	Cars
1	12
2	9
3	25
4	22
5	35
6	14
7	12
8	9
9	25
10	11
11	22
12	39
13	33

Week over week

The previous week had 7 days, this one 6. Compare trains and cars, not version counts.

This week against the previous week
Metric	Previous	This week	Change
Releases (trains)	13	19	+46%
Units of work (cars)	100	136	+36%
Pull requests merged	142	158	+11%
Issues closed	172	207	+20%
Real defects caught before merge	337	298	−12%
Cloud test runs, green	333, 93%	470, 89%	+41% runs
Automated tests	5,822	6,907	+19%
Line and branch coverage	87.32%, 81.47%	88.16%, 82.50%	+0.8, +1.0 pts
Lines added to main	112,276	93,155	−17%
Cash spent	≈ $92	≈ $137	a third reviewer
Google Cloud, on credits	$311.63	$347.28	+11%
Human-equivalent hours	4,020	4,134	+3%

Unit costs fall

One release per train instead of per car, parallel lanes, prompt caching, and review prompts narrowed to P1/P2.

Cost per unit of resultIndex, previous week = 100. Lower is better

Tokens per car fell from 96 to 53 million; list-price dollars per car from $38.46 to $21.24; per hour of equivalent work from $0.96 to $0.70.

Show as table

Cost per unit of result
Unit	This week
Tokens per car	56
Tokens per 1,000 lines	91
Tokens per hour of work	74
$ per car	55
$ per 1,000 lines	91
$ per hour of work	73

Against the industry

The industry’s own definitions and published thresholds. Results are compared, not the method: these benchmarks describe people who use AI tools.

DORA four keys (DORA 2024 thresholds)
Metric	This week	Elite	Verdict
Deployment frequency	3.3 releases a day	On demand, several a day	Elite
Lead time for changes	Median 3.3 h from PR to production	Under a day	Elite
Change failure rate	1 of 20 releases (5%)	5%	Elite, no real users yet
Time to restore	4 h 26 min	Under an hour (High: under a day)	High

Engineering benchmarks (LinearB 2026, 8.1 million PRs)
Metric	This week	Elite or industry	Verdict
Cycle time	Median 3.3 h	Under 25 h	Elite
Coding time	Median 24 min	Under 54 min	Elite
Pickup and review	Median 20 min	Pickup under 1 h, review under 3 h	Elite
PR size	Median 570 lines	Under 105	Needs attention
Merge frequency	139 car PRs in 6 days, one person	Over 2.2 per developer a week	About 63 elite developers
PRs closed without merging	2 of 489 (0.4%)	AI-written PRs merged within 30 days: 32.7%	Far above

Human equivalent

How much human work the week holds, by three independent methods. The honest answer is a range, and the most cautious method comes first.

One week of the factory, in weeks of a five-person teamTeam: 5 × 40 hours at €30 an hour, €6,000 a week

Lower estimate (QSM ÷ 2.6)12.9 weeks
Factory formula20.7 weeks
Check: PR throughput at LinearB elite21.0 weeks
QSM, unadjusted33.6 weeks

Lower estimate €77,600; factory formula €124,000; QSM unadjusted €201,700. The week actually cost about €122 in cash, €430 with cloud credits, and €2,570 at full list prices. Test code is left out of the QSM methods.

Show as table

One week of the factory, in weeks of a five-person team
Item	Team-weeks
Lower estimate (QSM ÷ 2.6)	12.9 weeks
Factory formula	20.7 weeks
Check: PR throughput at LinearB elite	21.0 weeks
QSM, unadjusted	33.6 weeks

Costs, three ways

Cash that left the account, cloud credits spent but not paid, and the list price a company without either would pay.

Cash
Train-wide reviewer, API	$59.60
Third car reviewer, API	$31.26
Agent subscription, weekly share	$46.15
Total	≈ $137

Google Cloud, on credits
Vertex AI (car review and other AI calls)	$194.99
Cloud Build (tests on 32 vCPU, images)	$149.08
Cloud Run, Cloud Storage, Artifact Registry, BigQuery	$3.21
Total	$347.28

Everything at list price
Agent sessions	$2,355
Gate reviewer	$95.81
Google Cloud	$347.28
Other reviewers	$90.86
Total	≈ $2,889

Left to pay after credits: $0.00. 98% of all agent tokens were cache reads. The limit this week was the subscription window, not money.

Codebase

Lines of code without blanks and comments. Tests now make up about half of all code.

Code and testsLines, counted with cloc

Code
Tests

Tests: 96,256 of 194,973 lines (49%), about one line of tests for every line of code.

Show as table

Code and tests
Snapshot	Code	Tests	Total
Two weeks ago	62,370	40,888	103,258
Last week	80,372	73,010	153,382
Now	98,717	96,256	194,973

Risks it names

A report that only shows wins is an advertisement. This one lists what could go wrong.

No real users yet: Stability in production is not proven. "Zero user incidents" is true and proves little.
One machine, one subscription: Both hit their limit this week. Growth needs a second window or paid API: a business decision.
Flaky tests: 89% of cloud runs green. It costs queue time, not correctness, but it erodes trust in a red build.
Large pull requests: Median 570 changed lines, the weakest number against industry benchmarks.
Credits run out: About $350 a week of cloud credits. The remaining grant needs confirming.
Self-measured: The numbers come from the factory’s own tools. They are reproducible, but not yet externally audited.

Where the numbers come from

Queries and scripts are reproducible from the journal views. Industry figures come from the publishers’ own documents.

Sources

Journal: BigQuery views for cars, trains, the train scorecard, review runs, reviewers by day, test runs and costs.
Billing: Google Cloud billing export, converted at the export’s own rate. The last two days are not final yet.
Code: GitHub API and git history of the main branch; lines counted with cloc on snapshots, without lock and generated files.
Tokens: Local agent session transcripts, de-duplicated by message, priced at the provider’s public list.

Definitions

Car: One task and one pull request.
Train: A set of cars released together.
Real finding: A reviewer finding checked against the code by the session running the train.
Unique: No other reviewer raised it on the same commit.
P0/P1: Would have harmed a user or their data, or blocked a release.
Change failure: A release that needed a fix in production. Planned follow-up releases do not count.

Industry benchmarks

DORA 2024 and 2025, State of DevOps
LinearB 2026 Software Engineering Benchmarks
QSM, typical software project (13,000+ projects)
COCOMO II Model Definition Manual
Capers Jones, Software Defect Removal Efficiency
Google Testing Blog, Code Coverage Best Practices
SonarQube, Sonar way profile
SmartBear, code review rates
Bulldogjob IT Community Report 2025

Get on the waitlist Next: Control it your way