Choosing a model: Auto and explicit picks

Last updated 15 Sep 2026

Overview

ToRun does not tie you to one model. Every chat turn, image, video, voice session and workflow step is routed to a model from the platform catalog — and you either let Auto choose, or name the model yourself.

What Auto actually does

Auto is capability-first, not brand-first. For a given turn it:

  1. Works out what the turn needs — plain text, vision (you attached a picture), image generation, image editing, tool calling, audio, a long context.
  2. Keeps only the models that support all of those capabilities.
  3. Narrows that set to what your plan and your funding allow.
  4. Ranks what is left by the published rate for the units the call will actually consume, and takes the best value.

Because the requirement is worked out per turn, one conversation can run on different models from one message to the next. Attaching a photo moves that turn onto a model that can see; asking for a change to an image you just made moves it onto one that can edit.

Picking a model yourself

The model picker on chat, image, video and workflow steps accepts an explicit model, and an explicit choice is honoured — the ranking never overrides it and nothing cheaper is quietly substituted.

The trade is that an explicit pick can be refused for a job it cannot do. Asking a text-only model to edit an image returns an error naming the mismatch instead of pretending.

When a provider fails

If a provider fails before it has produced anything — an outage, a rate limit, or a reply that comes back empty — an Auto turn is rerouted to a sibling model and answered. An explicit pick is not rerouted: you asked for that model, so you get its honest error rather than a quiet substitution. Deep Research is never rerouted, because its engine is part of what you chose.

What it costs

Every call is billed at the published rate for the model and the unit it used, and lands as one entry in your ledger with that rate stamped on it. Auto choosing a different model therefore changes what a turn costs — which is exactly why the last step reads the price list instead of following a fixed favourite.