Skip to content
In training

Nex

Our own model, learning where every request is worth sending. It is not finished, and this page says which parts of it are.

Why a model at all

The Gateway’s job is to put a request in front of the right model, and right now that decision is made by rules somebody wrote down. Rules are a reasonable start and a poor finish: they cannot tell a hard question from a long one, and they never learn from being wrong.

Meanwhile every call that goes through leaves a record of what a model was asked, what it produced and what that cost — which is exactly the data the decision should have been learning from all along.

Nex is that decision, learned rather than typed, and then the model behind it.

Where it actually is

One of these runs today. The other two are described in the future tense because that is what they are.

  1. 01Running now

    The learned router

    Every request through the Gateway already produces a row: what was asked for, what served it, how many tokens it took and what it cost. That is a labelled dataset about model choice, collected as a side effect of doing the job. The router learns from it — given a request, which model is this actually worth — and the honest version of that question includes the ones where the cheap model was fine.

  2. 02Being built

    Routing you can audit

    A router that picks for you is only worth having if it can show its work. Each routed call will carry the alternative it passed over and what that would have cost, in the same receipt the request already returns — so the saving is a number you can check rather than a claim on a pricing page.

  3. 03Planned

    The model itself

    A model trained by us, served through the same endpoint and priced on the same receipt as everything else the Gateway routes to. It is not trained, it has no benchmark to quote, and there is no date here because we would only have to move it.

Nothing here is something to wait for

The Gateway routes to every model it already supports today, on one key, with a priced receipt on every call. The routing data that trains Nex is a by-product of that, not a separate programme you have to opt into.