Eight minutes to run, twice a day. What it costs you to own for a month:
- CI compute, 480 minutes at ~$0.01/min, $4.80
- Engineering, one hour. Writing it, fixing the flake, re-reading the failure that turned out to be a timing issue. At a loaded $100/hr, $100
- Tokens, if an agent authored it, $10
~$115 a month. For one test.
Compute is 4% of that, and it’s the only line anyone optimizes. Shard the suite. Parallelize the runners. Shave the container spin-up. The other 96% is a person maintaining brittle selectors.
We charge somewhere between $8 and $20 per maintained test per month. No CI bill, it runs on our infrastructure. The reason we can do this so cheaply is that we have an amazingly good harness, and our humans-in-the-loop are trained to use that harness in the optimal way.
$20 versus $115.
Now look at which of those two numbers is larger. The review time costs more than the product. That’s the line I’d want a skeptic to sit with, because it means the remaining cost isn’t tooling, it’s attention, and there is no version of QA where you don’t pay that.
But the savings is the boring half of this.
At $115 a test, a hundred tests is $11.5k a month and you spend your planning meetings arguing about which hundred. At $18 all-in, that same budget holds six hundred.
That’s the number that matters. Not the discount. The coverage.
The three checkout variants. The flow only enterprise customers hit. The edge case that bit you in March and got a Jira ticket nobody picked up. None of it was blocked by budget. It was blocked by how much breakage your team could absorb in a sprint.
We price per maintained test because maintenance was the expensive part. Everything else was rounding.
