Switch the model for a tab

Agents: claude, codex, opencode, cursor · On: phone, desktop

Each tab runs one model. Change it by tapping the model button under the composer, or by typing /model.

Why you would use it

The tab was opened for a quick question on the cheap model and has turned into real work. Or the reverse: you are on the expensive model and about to ask something trivial. The model is per tab, so one tab can grind on a big refactor while another answers quick questions.

How to use it

Tap the model button (#modelpick) in the row under the composer, between the agent's name and the effort. It shows the model this tab runs, or auto.
The menu opens as a list of models, headed model for claude · now: Opus 5.5, with a tick on the one this tab is already using. Tap another.

Or, by typing:

Type /model. The menu becomes the same list of models as soon as the word is finished — no space needed — and keeps filtering as you type a value. It works at the end of a message already written: picking a model there sets it and takes /model out of the draft, leaving the message.
Send /model on its own to see what the tab is on instead.
Type /model opus and send it to set it directly. The product name works too: /model Opus 5.5 resolves to opus.

Both the picker and the typed command go through the same place, so they cannot mean different things.

What you see

The button's label is the model's product name, not a caption: Opus 5.5, GPT-5.6 Sol, or auto in a dimmed style when nothing is set.
Long ids are shortened for the label: claude-opus-5-5[1m] reads as Opus 5.5, a trailing snapshot date is dropped, and gpt-5.6-sol reads as GPT-5.6 Sol.
The model this tab is on carries a green tick in the menu, and its note starts with in use — in use · agentic coding and long-horizon work. It follows what the tab is really running: auto holds the tick only until the first turn, after which the agent reports the id it started with and that model's row takes it.
A point release is ticked on its family's row. Running claude-fable-5-1 ticks Fable 5, and the header reads now: Fable 5.1.
The tooltip is Model: opus, or Model: whatever the CLI is configured for when nothing is set.
/model with no argument answers with a toast: model: opus, or model: default.
A second tap on the button closes the menu.

Options and settings

claude (OPTION_VALUES in public/app.js):

Value	Shown as	Note
`opus`	Opus 5.5	agentic coding and long-horizon work
`sonnet`	Sonnet 5	near-Opus quality at Sonnet cost
`haiku`	Haiku 4.5	fastest and cheapest
`fable`	Fable 5.1	most capable for your hardest and longest-running tasks
`default`	auto	whatever the CLI is configured for

codex: gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark, and default. Their notes run from the most capable — complex, demanding work to ultra-fast coding.

opencode has no table. Its models come from the list of agents the server reports, because they depend on which providers it is logged into, plus default labelled whatever the agent is configured for.

Cursor's models also come from that list, using the model IDs and names advertised by ACP. The picker sends the exact ID ACP accepts (including any bracketed parameters); the single auto row uses ACP's default[] model.

Limits and known gaps

Tapping the ticked row is not a no-op. It sends the alias, so a tab running claude-opus-5-5[1m] that is ticked on Opus 5.5 restarts on plain opus and loses the 1M-context variant.
The setting is per tab, not global. A new tab opened with /new copies this tab's model.
The value sent to the CLI is the alias (opus), not the product name; the row's tooltip shows what will be sent.
On claude and opencode the model is a process flag, so the agent process is stopped and the next turn starts a new one. Nothing is lost: that start resumes the same vendor session.
A change made mid-turn is held. The process that owns the turn keeps it, and the restart happens before the next turn.
On codex the model rides on turn/start as an override, so it lands on the very next message with no restart.
opencode's catalogue depends on which providers it is logged into. If discovery has not finished, the list can be empty and the client retries within the server's 60-second budget.
The button is disabled with no tab open.
Choosing a model from the menu does not clear your half-written message. Only when the draft is the command (/model …) is the box emptied.