RouterPlus launches with pooled inference pricing for small teamsMachine translation

Inference pricing has a size penalty. The best per-token rates go to whoever can commit the most volume.

RouterPlus groups demand from many teams into one large deal with a provider, and each team pays that rate on its share of the traffic.

Large buyers have done this with cloud commits for years. Applied to tokens, I’d guess it works best for popular open models, where many teams already want the same weights served on the same kind of hardware.

For a small team, this is a route to bulk pricing that doesn’t require being a bulk buyer. They also test models on your real usage and pick the one that's good enough at the lowest cost. Many teams pay for a top model on every task when a cheaper one could do most of the work.