PostHog says decision models run parts of its product 444.6x cheaper
The numbers are precise to one decimal place and the post does not say what they were measured against.
PostHog published a write-up of how it uses decision models, saying models like Jev can do some of the work "up to 444.6x cheaper and 193.6x faster" than constant LLM calls. The framing is a cost argument first. In PostHog's words, building a product that requires constant LLM calls gets real expensive, real fast.
Building a product that requires constant LLM calls gets real expensive, real fast.
What the post does not say is the part that matters if you want to reproduce it. There is no named baseline model, no named workload, and no description of how the two systems were run against each other. A figure given to one decimal place implies a measurement, and the measurement is not shown. 444.6x and 193.6x are PostHog's own numbers from PostHog's own product paths.
What you can actually take from it
The directional claim is unremarkable and probably right. Swapping a general purpose LLM call for a smaller decision model on a narrow, high volume path is the standard way these cost curves get bent, and PostHog is saying it did that in production rather than in a notebook. That is a real signal about whether the pattern survives contact with a product people pay for.
The magnitude is a different thing. A ratio that large usually means the expensive side of the comparison was a frontier model being asked something a classifier could answer, on a path called often enough for the per-call difference to compound. Pick a cheaper baseline or a harder decision and the ratio collapses. Treat 444.6x as the top of a range someone measured once, not as what your workload will do.
Diogo Almeida, reacting to the post, said that now 50% of his most favourite software companies use Jev. That is a vibe check, not a data point, but it is the kind of thing that moves adoption anyway.
For people building on it, the honest comparison is the one you run yourself on your own path, against the specific model you would otherwise call. We have already covered a case where that comparison went the other way, with Jason Cohen unable to get Jev to match Sonnet on classification.
now 50% of my most favorite software companies use jev
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.

