Availability
Entry points
- Model button (bottom right of the composer): shows the Harness, the model vendor icon, and the model name. Small text after the model name shows the source: Subscription for your subscriptions, BYOK for an API key you added, and nothing for models included with Plus.
- Effort button (next to the model button): shows the current level, for example Medium or High · Ultracode. It is hidden when the model does not support adjusting effort.
- Context usage ring (in a session, to the left of the model button). See Context usage ring.
- For shortcuts, see Keyboard shortcuts.
- When plan limits are reached or an account is disconnected, Use another model in the dialog also opens the model picker. See Plan limits reached.
Model picker
The top part of the model picker is the Harness list (the selected one has a check mark), and the bottom part is Model, with two tabs, Subscriptions and BYOK. Models included with Plus and your own API keys are both in the BYOK tab.Supported Harnesses
The Harness list follows what the product shows. It currently includes Claude Code (the default Harness for new sessions, and the only one that supports Ultracode), Codex CLI (some models have an Ultra level), OpenCode, Hermes Agent, DeepSeek, and Pi. Whether you can add a message during a run or edit a sent message depends on the Harness. See Composer.Harness and model source combinations
The model picker lets you choose only models the current Harness can use. Other groups are grayed out with a reason.- Plus models: available models vary by Harness.
- Your subscriptions: each subscription maps to a specific Agent, for example Claude subscriptions are for Claude Code only and ChatGPT subscriptions are for Codex CLI. See Supported subscriptions.
- Your keys: Claude Code needs an Anthropic compatible API, and Codex CLI needs an OpenAI compatible API. See API format and Harness.
Choose a Harness and model
- On the new session page, click the model button at the bottom right of the composer.
- Click a Harness. After you change the Harness, the model returns to Default model.
- Switch to the Subscriptions or BYOK tab and click a model.
- If needed, click the effort button to adjust reasoning effort.
- Enter your message and send it. The session starts with this combination.
Default selection
- New sessions reuse the choice from the last session you created (Harness, model, source, and the reasoning effort you actively selected). The first time, it is Claude Code + Default model.
- If you never actively selected an effort, the effort button shows the model’s recommended level.
- If the remembered model or source is no longer available, it returns to Default model.
- Ultracode is off by default in new sessions.
- Switching models in a session affects only that session and does not change the defaults for new sessions.
- With several subscription accounts from the same provider, the account marked Default is used. See Default account.
Harness is fixed after the session starts
After the first message in a session is sent, the Harness cannot be changed. You can still change the model, source, and reasoning effort, and changes take effect from the next turn.- The Harness area of the model picker shows only the current Harness and a lock. Hover to see This task’s harness is fixed. Start a new task to use another.
- When the session belongs to a project, you can click New task with another harness › to create a new session in that project with another Harness.
Switch models in a session
- In the session, click the model button and click another model (you can switch across sources).
- If the current context fits in the new model, the switch happens directly and a Switched to model name row is left in the session. If it does not fit, you are asked to confirm compaction first. See Compact and switch.
- All earlier session history is kept and the new model continues from it.
- You cannot switch while the turn is running. It shows The task is running. Switch models after this turn ends.
- If the switch fails, the reason is shown. You can try another source or model, or retry.
- If the session is paused because plan limits were reached or an account was disconnected, it continues automatically after you switch models or accounts.
Compact and switch
When the current context does not fit in the target model, a Switch to model name? confirmation appears in the popover: the chat is compacted into a summary first, then the new model picks it up, and earlier details may be lost. When the effort will change, it also shows Effort after switching: X.- Cancel: returns to the model list and keeps the current model.
- Compact and switch: compacts and switches to the new model, and leaves a Compacted the chat and switched to model name row in the session.
- If the session is too long to compact, it shows This chat is too long to compact. Try starting a new session.
Context usage ring
In a session, the small ring to the left of the model button shows the current context usage. Hover to see the percentage used and the token count. Near the limit, the ring changes color and the session is compacted automatically.Reasoning effort
Reasoning effort determines how long the model thinks before answering: higher levels give more thorough answers but are slower and use more. Which levels each model has and which one is recommended depend on the “model + source + Harness” combination, as shown in the product. You can also change it in a session, and changes take effect from the next turn.Two styles
Level names
- Effort level list: lists only the levels the current model supports. The recommended level is marked Recommended. Click a level to apply it. The Ultracode toggle is below the list.
- Effort slider: drag or click to choose a level. The line below describes what the current level is for. Reset to default resets to the recommended level.
Ultra level
Some models offer an Ultra level in Codex CLI, at the right end of the slider. Selecting it opens a Turn on Ultra reasoning? confirmation, which notes that usage is about 2–3× that of Extra high and each task takes longer.- Click Turn on Ultra to apply it. Click Cancel to keep the original level.
- In a session, you can check Don’t ask again this session.
Ultracode toggle
Ultracode is Claude Code’s multi-agent parallel feature, which splits a task into multiple subtasks and completes them in parallel. Usage goes up noticeably. It is a separate toggle below the Effort level list and does not change the Effort level. When it is on, the effort button shows something like High · Ultracode. In a session, it takes effect from the next turn.- Whether it can be turned on depends on the Harness, cloud server, model, and source: it needs Claude Code, a ready cloud server, and a selected model that supports Extra high.
- When the conditions are not met, the toggle is grayed out with the reason. In a new session, if changing the model or source means the conditions are no longer met, the toggle turns off automatically.
How effort changes after switching models
Keyboard shortcuts
Without Plus
- The model picker opens as usual. In the BYOK tab, the Plus models group is grayed out and shows Subscribe to Plus to use these models. Click Get Plus to open the Plus subscription dialog.
- On the new session page, the message you send is held and an upgrade card is shown. After you subscribe, it is sent automatically. See Without Plus.
Notes and limitations
- To change the Harness, create a new session.
- Compaction loses earlier details. If you do not want to compact, click Cancel, or choose a model with a larger context limit.
- In a new session, clicking a model in an incompatible group switches the Harness automatically. Watch for the notice.
- The model list, levels, and recommended levels may change. Follow what the product shows.