Jev Daily

Mike Taylor priced Jev video labelling at 14 cents an hour

One line item in a log of 65 things shipped while burning roughly 14 billion tokens over four weeks.

Mike Taylor ran a Twitch clip labelling test to see whether Jev could label an hour of video now that it reads images. His number is 14 cents an hour, against about $6 for what he calls the 2024 pipeline, and he reports it scored the same as a route costing more than twice as much.

That is roughly a fortyfold drop in cost on one labelling task, with the quality claim resting on a comparison against a pricier route he does not name. The entry appears in a log of 65 things he shipped while spending about 14 billion tokens testing Fable 5.2 over four weeks, so it is a line item rather than a write-up with a methodology section attached.

Mike Taylor
@hammer_mt
X
14 cents an hour, against about $6 for the 2024 pipeline, and it scored the same as a route costing more than twice as much.
Nov 9, 2026 · View on X

What the number is and is not

One task, one person's pipeline, measured against his own older setup. There is no published benchmark here, no sample size for the clips, and no breakdown of what "scored the same" means in labels or accuracy. Treat it as a cost datapoint from someone who actually ran it, which beats a vendor estimate, and nothing more.

Taylor is unusually blunt about the limits of his own measurement setup elsewhere in the same log. Writing up a case for moving benchmark runs to the cloud, he notes that every run inherits one laptop's quirks, that there is no clean lab, and that models can peek at the eval files. That caution applies to his own numbers as much as anyone's.

For people building on it

If you have a video labelling job currently sitting on a 2024-era pipeline, this is a reason to re-price it, not a reason to assume the ratio holds for your clips. Image reading is new behaviour, and the cost advantage was measured on Twitch footage with whatever label schema Taylor happened to need. The cheap part is the per-hour inference. The part nobody has priced for you is finding out whether the labels are good enough for your downstream job, which in Taylor's own framing is the work that still needs a person looking at outputs.

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.