How models are chosen

A routing engine that learns from your company's own work, and keeps learning with every conversation.

The starting point

Each request starts from a default model: the one your company set as its default on the Models page, or, if none is set, the model the request names. Sovereign moves off it only when the move earns it:

• to a cheaper model when that model will do the work as well;
• to a stronger model when the work needs more than the default reliably gives.

Reasoning effort starts the same way, from the level your tool asked for, or the model's own default.

The Saved column in the dashboard is always measured against the model the request asked for. Spending more on a stronger model shows as an upgrade, not as a loss.

A decision model built to route

At the heart of Sovereign is a purpose-built decision model. It reads every request in context: what the work is, where it sits in its conversation, and what your company has learned about work like it. It weighs quality against cost across every model you can use, and returns its choice, with its confidence, in a fraction of a second.

The Sovereign score

Public benchmarks say how models do on someone else's tests. The Sovereign score says how they do on yours.

• Frontier-model evaluation. Frontier models evaluate responses from your real traffic, blind to which model wrote them, and score each model out of 10 on the work it actually does for you.
• Online, continual learning. Every evaluation updates the scores as your traffic flows, for each kind of work your company does (a support bot, contract review, coding) and across the company as a whole. Recent results count for more, so the scores keep pace as models improve and your prompts change.
• Uncertainty-driven. Evaluation is busiest where Sovereign is least certain and eases off as its confidence grows, so it learns fast at first and stays quiet once it knows.
• Thompson sampling. New models earn their place: each gets a share of real work in proportion to the chance that it's the best one for it, never at a higher price than your default, until Sovereign knows where it shines.

Each model's Sovereign score is on the Models page, once it rests on enough of your work to mean something.

Conversations stay together

A conversation stays with the provider it started on, so its prompt cache keeps paying off and the model can read its own earlier reasoning. Sovereign changes course mid-conversation only when the change is clearly worth it.

Models that can't serve a request are never chosen

If no model your company allows can serve a request, it fails with a 422.

Which model answered

Every response carries an x-sov-model header with the model that answered, as listed by GET /v1/models.

Model names

Use any id from GET /v1/models, or the OpenRouter name for the same model. Names Sovereign doesn't know are still answered, but there is nothing to measure savings against.

Your company's choices

The models you can use are the ones your company's provider keys reach. On the Models page an admin can add more of them, remove the ones your company shouldn't use, and set a default model for everyone.