Hello Guys
I Want to ask about the work auto-model do and how cursor make the auto model routing the task to specific model and in the next step choses another model, is there any docs i can read?
Hey, thanks for the question. We don’t publicly share the exact Auto routing algorithm, but the general idea is that Auto picks a model from the ones available for your request based on the task type, context, and current model availability and load. Within a single session, the model can change from step to step. That’s expected and it’s meant to keep a reasonable balance between quality and speed.
Auto and Composer use a separate usage pool that’s different from the API pool, where billing follows the rates of the specific model. More details about pools and pricing: Models & Pricing | Cursor Docs
If you want a fixed model, you can pick one manually in the model picker, and then Auto won’t override it.
Hi. About where does Auto mode sit on Cursor bench? I know it uses many models, but a general idea… Is it like Sonnet 4.6? Grok 4.3? Opus 4.5? Thanks.
Hey, Auto doesn’t map to a single model on any benchmark that you can pin it to. Within one request, and often even within one step, it might use different models under the hood depending on the task type, context, and current availability or load. So it’s not like Auto equals Sonnet 4.6 or Auto equals Grok 4.3. The mix and the proportions change.
The goal of Auto is to keep a reasonable balance of quality and speed by using the available pool of models, not to match one specific model. If you need a predictable level and stable behavior for a specific workflow, it’s better to pick a model manually in the picker, then Auto won’t override it.