Mixture-of-experts pricing was wrong for four months
We were counting tokens against total parameters rather than active experts. Everyone affected has been credited; nobody had to ask.
A mixture-of-experts model has a large parameter count and a much smaller number of parameters active per token. Mixtral 8x7B is 46.7 billion parameters, of which about 12.9 billion run for any given token. DeepSeek V3 is 671 billion, of which roughly 37 billion run.
Our cost model priced against the total. That is the right number for how much memory a replica needs and the wrong number for how much compute a request consumes, and we used it for both. The result overcharged MoE requests by about eight per cent.
The credit
Every affected account was credited automatically on 6 May, backdated to the first request. Nobody filed anything and nobody had to notice. If your invoice for May looks strange, that is why, and the line is labelled.
The lesson, such as it is
A cost model that quietly reuses one number for two purposes will be wrong on the day those purposes diverge. Memory footprint and active compute were the same number for every dense model we had ever served, so nothing forced them apart until a MoE model arrived and nothing broke loudly enough to notice.