Each AI feature in your project writes with its own model. Every feature starts on Fast. You can pick a bigger model for one feature, say AI answers, without changing the others. Owners and admins can change this.

Pick a model

  1. Open Settings → Models. The page is called AI models.

  2. Find the feature:

    Feature

    What its model writes

    AI answers

    Answers to questions in your docs, on the portal and in the widget.

    Ask a person (summaries)

    The summary sent with a question passed to your team.

    Support inbox (drafts and summaries)

    Replies drafted from your docs and summaries of conversations, in your Inbox.

    Changelog drafts and edits

    Draft with AI and Edit with AI on changelog entries.

    Autopilot decides

    Reads each merged pull request, its code and your help center, and decides which articles to fix and which new ones you need.

    Autopilot double-checks

    When Autopilot suggests a new article, decides again from scratch, so fewer duplicates are written. It can be Off.

    Autopilot writes

    Each fix and new article, the summary, the changelog line and updates on posts. Verify uses it too.

  3. Choose one of its cards.

  4. Choose Save. "AI models saved" confirms it.

The AI answers feature with Fast picked, the Balanced card, and Custom model, each with roughly what one answer costs

The choices

  • Fast: the default, and what a feature uses until you change it. It's Tuti's default model for that feature, so if Tuti changes the default, the feature follows. Choosing Fast again undoes an earlier pick.

  • Balanced: a larger model that can write better. It costs more than Fast.

  • Best: the largest model offered. It costs the most.

  • Custom model: any other model id. A box for the id, like provider/model, appears under the cards. As you type, it suggests the models available, with their prices per million tokens.

What each choice costs

Each card shows the model it uses and roughly what one use costs in AI credits, like "About 0.85 credits per answer". A use is an answer, a summary, a reply, a draft, or for Autopilot a merged pull request. The estimate is for a typical use: a long question or a large pull request costs more.

Autopilot's three rows each have their own estimate: deciding and writing per merged pull request, and the double-check per double-check, which only runs when Autopilot suggests a new article. A merged pull request costs the steps that run for it, added up. Only models that can use tools are offered for them. More on these, and on Autopilot's Advanced tuning on Business, is in Autopilot's models and Advanced tuning.

What each feature really spent is in Usage, at the bottom of Settings → Billing. Autopilot's three steps and Verify show together under Autopilot. See AI credits.

If saving doesn't work

  • The model id isn't known: the page says there's no such model. Check the id, or pick it from the suggestions.

  • "Couldn't check the models just now. Try again in a moment.": save again in a moment.

  • The model can't call tools: one of Autopilot's rows was given a custom model that can't use tools, which Autopilot needs. Pick one the row suggests.

When saving fails, none of the page's changes are saved.