Reduce token spend by 90%
Task Complexity and model capability routing to improve AI performance and reduce cost.
Capabilities
No single model is best at everything.
Every model is trained on different data toward different goals, so each one ends up with its own strengths. One is strongest at multi-step reasoning, another reads long documents more carefully, another answers routine questions well for a fraction of the price. Sending every request to one model means paying frontier prices for work a smaller model handles just as well, and missing the model that would have done the hard work best.
Integration
It works like the API you already call.
Sansa is an OpenAI-compatible chat completions API. Point the client you already use at Sansa and set the model to sansa-auto. Requests go in with the same shape and responses come back with the same shape, so the rest of your code stays exactly as it is. Everything that happens next happens inside that one call.
Prediction
It reads what each request needs.
When a request arrives, Sansa runs it through an encoder-only transformer trained on a capability space shared by prompts and models. The same axes describe what a task demands and what a model can do. In a single pass, the encoder predicts where the request sits on many of those axes at once, like reasoning depth, coding, math, domain knowledge, context length, and tool use.
Selection
Then it finds the best model for the price.
Because models live in the same space, Sansa can place the request’s needs against every model in the pool. It estimates how well each one would answer and what it would cost, then keeps the Pareto-optimal options, the models that no other model beats on both quality and price. From that frontier it picks the cheapest model that meets what the task needs.
Delivery
Your request goes to a stable provider.
Sansa proxies the request to a reliable provider for the chosen model. Within a thread, follow-up turns stay with the model and provider that already hold the conversation in their prompt cache. You are never routed away from a warm cache, so long conversations keep their fast, discounted cached tokens instead of paying to resend context the provider has already seen.
Results
Higher quality at a lower total cost.
You get the model’s response back like any other completion. Across your traffic, hard requests still reach frontier models and routine ones go to models that cost a fraction as much. The mix is what pays off. On MMLU-Pro, Sansa Auto scores 93.3% at $4 per million output tokens, above every individual model tested, including ones that cost six times more.
More AI. For Less.