Dark Code Factory Prices User Stories at $10.88 Each, Burning 7.7 Million Tokens

1 articles · Updated · O'Reilly Media · Sep 2

$837.53 in API-equivalent spend shipped 77 stories in the measured repository sample, putting the autonomous SDLC factory at 7.7 million tokens and $10.88 per delivered story.
595.7 million tokens across 374 stage attempts included failed runs, retries, bugfix loops and five failed stories, while per-story costs ranged from $3.02 to $43.24 with a $9.56 median.
95.4% of tokens were cache reads, making the system "mostly a reading machine"; cache traffic accounted for 77% of cost, far outweighing fresh input at 1.6% and shifting optimization toward cache management.
Session logs, not the factory ledger, became the ground truth after an audit found the meter undercounted spend by about one-sixth—$694.65 versus $837.53—because 57 attempts were overwritten or never recorded.
The author says the real cash outlay was a $200-a-month subscription rather than per-call billing, implying more than $1,088 in monthly API-equivalent usage and highlighting how flat-fee quotas can mask true software-production costs.

Sources

If cache management is the true hidden cost of AI coding, how long until flat-rate subscriptions disappear entirely?

Does an AI coding factory actually write better software, or does it just brute-force millions of tokens until something finally compiles?

When an autonomous AI factory fixes its own tracking bugs, who is truly in control of the software development lifecycle?

1 total