# Harnesses and models

The Claude Code and Codex versions the Companion runs, and every model Cabane offers with its tuning controls.

Last updated 9 November 2026

A harness is the program that runs an agent's turns on their device: Claude Code, Codex or OpenCode. For Claude Code and Codex, Cabane offers a fixed list of models, each with its own tuning controls. For OpenCode, it lists the models your OpenCode server reports.

The tables on this page come from Cabane's own model list and Companion release, so they match this version of Cabane. To choose a model for an agent, see [Model and tuning](https://cabane.ai/docs/model-and-tuning).

## What the Companion runs

The Companion includes its own Claude Code and Codex. Agents' turns run on these copies, not on a Claude Code or Codex installed separately on the device.

Companion 0.6.233 and later include:

| Harness | Version | Package |
| --- | --- | --- |
| Claude Code | 2.1.319 | Claude Agent SDK 0.3.319 |
| Codex | 0.167.0 | Codex SDK 0.167.0 |

A device on a Companion older than 0.6.219, the oldest one Cabane runs, shows **Update required** beside its version.

To see which Companion a device runs, open the workspace's **Settings → Devices** and choose the device. Its page shows "Companion v…".

## Models

Models are listed in the order the **Model** menu shows them. Each tuning control lists its values, and bold marks the value the model starts on.

### Claude Code

| Model | Thinking | Effort |
| --- | --- | --- |
| Fable 5.1 | **Adaptive**, Extended, Off | Low, Medium, **High**, Extra high, Max |
| Opus 5.5 | **Adaptive** | Low, **Medium**, High, Extra high, Max |
| Sonnet 5.5 | **Adaptive** | Low, Medium, **High**, Extra high, Max |
| Haiku 5.5 | **Adaptive**, Extended, Off | Low, **Medium**, High, Extra high, Max |
| Opus 4.8 | **Adaptive**, Extended, Off | Low, Medium, **High**, Extra high, Max |

### Codex

| Model | Reasoning |
| --- | --- |
| GPT-6 Astra | Low, **Medium**, High, Extra high, Max, Ultra |
| GPT-6.1 Sol | **Low**, Medium, High, Extra high, Max, Ultra |
| GPT-6.1 Terra | **Low**, Medium, High, Extra high, Max, Ultra |
| GPT-6 Luna | Low, **Medium**, High, Extra high, Max |

### OpenCode

The Companion doesn't include OpenCode. You install it and keep `opencode serve` running on the device, and Cabane lists the models that server reports, so the list depends on the device and its providers. OpenCode models have no tuning controls in Cabane.

## Retired models

When Cabane replaces a model, the old one leaves these tables and the **Model** menu. An agent already set to it keeps it: Cabane still sends that model to the harness, and whether it runs is up to the model's provider. If the workspace used the model recently, the **Model** menu still lists it at the end, by its id, such as `anthropic/claude-sonnet-5`.

## Staying current

Newer harness versions arrive with Companion updates. If a model needs a newer Companion than the oldest one Cabane runs, its table says which.

With automatic updates enabled for the Companion's service, it updates itself and can restart during a turn. The device's page in **Your settings → Devices** shows "Auto-update on" and the last update, if any. Without the service, run `cabane update` on the device when no agent is working, including on a device that shows **Update required**. The setup guides cover both: [Mac setup](https://cabane.ai/docs/mac-setup#3-keep-it-running-and-update), [Windows setup](https://cabane.ai/docs/windows-setup#3-keep-it-running-and-update) and [Server and VPS setup](https://cabane.ai/docs/server-and-vps-setup#3-keep-it-running-and-update).