A simple question should not cost what a hard one costs. The Gateway tells them apart with a model we run, and the decision itself never leaves.
The judgement is made by a model we operate — on every plan, not only the expensive one.
AWS and Azure both bill routing as a line of its own. Here it runs in the same pass as the security checks you already pay for.
Classification comes from a public leaderboard, not from our own scoring.
A classifier we operate sorts the request into one of five classes — in our own EU infrastructure, or on your hardware in Enterprise on-premise.
The class decides which models are eligible to answer it.
Of those, the cheapest one that can handle it.
No part of the request goes to a third party in order to be judged.
The strongest wording a hosted vendor uses is “we do not store it” — never “we do not send it”.
On our cloud the judgement is made by a model we run in Europe, never by the model vendor you are sending the request to. On your own hardware, in Enterprise on-premise, the same model runs where nothing leaves at all. Either way there is no surcharge, because it runs in one pass with the security checks you are already paying for.
Competitor prices as published in September 2026.
We do not guess which model belongs in which class — a script derives it from a public leaderboard and can be regenerated at any time. On 18 of 45 models it corrected our own manual classification, in both directions.
And the uncomfortable part: on RouterBench, a public benchmark, reading the request against a fixed threshold gained nothing at all. The reason is in the data — for 87% of those tasks there is no cheaper model that would have sufficed. So we do not promise a percentage saved. We offer to measure it.
A provider outage, an exhausted quota or a rate limit, and the request is taken over by another provider in the same class or above — only from models that application is allowed to use. The answer then carries a line naming the model that wrote it. Switchable off for an application that would rather have the error.
How much of your traffic a cheap model can carry depends on what your people do. A competitor cannot work that out for you — they would need your prompts on their servers. We measure it on yours.
After a month you get your own number: what share of traffic, what that is in money, and where routing does not help. If it turns out not to help, we will say so.
The measurement is a screen in the console, not a service we perform: it takes your stored requests, has both a cheap and an expensive model answer them, and you judge which answer would have done. The sides are labelled A and B — anyone who knows which is which decides by expectation, and then the number is not yours. You get the result with a spread, because a share of thirty judgements is a range, not a figure.