Hi, I'm Hisham, and I built SentioGate. I kept seeing the same pattern: most requests didn't need the model they were sent to. Simple questions went to expensive frontier models, and switching by hand across providers and keys was tedious and easy to get wrong. So I built a router that uses your own keys, adds no markup, and only picks from models you've approved in your vault. For each request it classifies the prompt, then picks the cheapest model that's good enough, based on live prices and measured success rates and latency. The approach changed a lot while I was building it. Early on, a single rate-limit error could knock a good model out of rotation, so I moved to smoothed success rates and quality bands. I also built a "cascade" mode (try a cheap model first, escalate if needed) and answer-length caps. In my tests neither saved money, so both ship turned off by default. What actually made the difference was caching, passing provider prompt caching through, and simply picking well. For the launch I ran a head-to-head on the same CommandCode key with 60 hard prompts. Our auto got 59/60 at about 62% lower estimated cost than a dedicated decision model picking the model, and about half the median latency. That's a small sample, and part of the gap is the decision model's own call cost, so I'd treat it as encouraging rather than proof. I'm happy to publish the prompts and the test code. I'd love feedback, especially on how it routes and on where it gets the choice wrong.

Zero-markup OpenAI-compatible router.

SentioGate — Zero-markup, BYOK intelligent model router.

SentioGate — Zero-markup, BYOK intelligent model router.