> ## Documentation Index
> Fetch the complete documentation index at: https://docs.usehenry.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Choosing a model

> Presets, the full catalog, reasoning levels, working style, fast replies, and per-task models.

Henry is model-agnostic. Admins choose the model and reasoning depth for the workspace in **Settings → Models**, and can override it per task. Memory lives in Henry, not in the model, so switching models keeps everything Henry has learned.

Henry uses the model you picked, full stop. There is no silent failover to a different model.

## Presets

Three presets cover most teams. Each card shows the model, its reasoning level, an estimated cost, and our take.

| Preset                  | Model                             | Our take                                                                                                                                                 |
| ----------------------- | --------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Ultra** (recommended) | Claude Fable 5, medium reasoning  | Understands what you actually mean with the least hand-holding, and stays sharp on the longest, most complex work. The closest thing to a real teammate. |
| **Smart**               | Claude Opus 5, medium reasoning   | Handles most day-to-day work well. Very complex or ambiguous asks land better on Ultra.                                                                  |
| **Balanced**            | Claude Sonnet 5, medium reasoning | A lot cheaper per turn. The right pick when Henry is a workhorse for high-volume, lower-stakes tasks.                                                    |

A workspace that has not chosen runs on **Smart**.

## All models

Below the presets, **All models** lists the full catalog with a relative cost label for each (base rate, cheaper by a percentage, or a multiple of base). The catalog includes models from Anthropic, OpenAI, xAI, and Moonshot; some entries are served by US inference providers and are marked as such.

## Reasoning levels

Each model accepts a range of reasoning levels: **None**, **Low**, **Medium**, **High**, **Extra high**, and **Max**. Higher levels think longer and cost more per turn. Set the level alongside the model. A level a model does not support is rounded to the nearest one it does.

## Working style

**Working style** sets how deeply Henry works on each conversation across Slack, Teams, and the web app:

* **Quick**: fast answers with a small tool budget. Best for simple lookups. Henry says when a question deserves a deeper pass.
* **Standard**: the balanced default, with enough tool calls and thinking for most day-to-day work.
* **Deep**: maximum thoroughness. Henry cross-checks, paginates, and uses many tool calls. Slower and costs more per turn.

In the web app, the composer's mode pill lets you pick Quick or Standard for a single conversation.

## Fast replies

Add `!fast` anywhere in a Slack message to answer that one message with a cheaper, lower-reasoning model. Choose which model handles fast replies under **Fast replies**, or leave it on the default.

## Per-task models

Each scheduled task can pin its own model and reasoning level in the task dialog. The default follows the workspace setting. Putting mechanical tasks on **Balanced** is the easiest way to cut recurring cost. See [Creating tasks](/user-guide/tasks/creating-tasks).

## Your own model API key

Enterprise workspaces can run Henry on their own Anthropic or OpenAI key. See [Bring your own model key](/developer-guide/bring-your-own-model-key).

## Parallel helpers

For large read-only research, Henry can split the work across parallel helpers that each handle a piece and report back. Henry verifies their findings before answering. Helpers cannot write or stage approvals.