Model Selector

Use the model selector to choose the model a thread runs on and the effort the model works at. The selector also shows who pays for each model. If you do not want to choose, leave it on Staff Pick.

You pick the model, and Amp builds the agent around it. Each model in the selector has its own system prompt and tools, and both change when you change the model.

Pick a Model

In the web app, open a new thread and select the model button below the prompt box, or press Ctrl+S. Staff Pick is first in the list, followed by your Top Models. The card next to the list describes the highlighted model and names the connection that serves it.

Three more rows are below the list:

Effort sets how much the model thinks before it answers. See Effort.
Speed switches between Standard and Fast, and Plaid for the models that offer it.
Other Models lists every other model and agent you can use. Type to search the list. Your own agents come first, and the models follow, grouped by provider.

The selector stays open after you pick a model, so you can set the effort next. While it is open, ← and → change the effort.

In the CLI, press Ctrl+S to open the same list. Use ↑ and ↓ to choose a model and ← and → to change the effort. Type to search every model, or choose Other models to see the full list.

Staff Pick

Staff Pick is the model the Amp team would choose for most work. Today it is Claude Opus 5.5. We keep testing models, and we change the pick when another model does better.

New threads start on Staff Pick unless you change New Threads Start On. A thread that has already started keeps its model when the pick changes.

Effort

Effort is how much the model reasons before it answers. A higher effort takes longer and uses more tokens. A lower effort is faster and costs less.

The selector offers only the effort levels the model supports and marks the recommended one. Each model starts at its suggested effort, which you can change in Top Models. An effort you pick yourself stays when you pick another model. If that model does not support the level, Amp uses the closest level it supports.

Switch Models in a Thread

You can change the model and the effort of a thread at any time. Open the selector from the model button in the thread’s prompt box, make the change, and send your next message. The new model answers that message, and the transcript records the change, e.g., “Switched to GPT-6.1 Sol · High”.

In the CLI, press Ctrl+S in the thread and pick a model. Only the thread’s owner can switch.

A switch has a cost. A prompt cache belongs to one model, so the new model reads the whole thread without a cache, and the first message after a switch costs more than usual.

A thread that runs an agent from a plugin or Claude Code keeps that agent.

New Threads Start On

Open Settings → Model Selector and choose what a new thread starts on under New Threads Start On:

Staff Pick starts every new thread on the Staff Pick. This is the default.
Last Choice starts a new thread on the model you picked last.
A fixed model starts every new thread on the one model you choose from the list.

The setting chooses only where a new thread starts. You can pick another model in the selector before you send the first message, and the model you pick for a thread always wins.

If a fixed model is no longer available to you, new threads start on Staff Pick.

Top Models

Top Models are the models listed at the top of the selector, below Staff Pick. Every other model stays under Other Models. Amp suggests a list to start with, and the list follows our suggestions until you change it.

In Settings → Model Selector, under Top Models, you can:

Press Add Model to add a model. You can search by name. The list holds up to 8 models.
Drag a model to change its place in the list.
Set the Suggested Effort of a model, which is the effort the model starts at when you pick it.
Remove a model from the list. You can still pick it under Other Models.
Press Reset to Suggested to go back to Amp’s list and order.

The badge on each model shows who pays for it, as the selector does.

Who Pays

The icon on each row of the selector shows what a model bills to. It is one of your subscriptions, one of your API keys, or Amp credits.

Model Routing decides which connection serves a model. Your personal connections come first, then your workspace’s connections, then Amp. To change who pays for a model, add a connection or change the order in Settings → Model Routing.

A model that only one of your connections serves, and that Amp does not offer itself, also appears under Other Models. See Models Amp Does Not Know.

Your Agents

The first group under Other Models is Your Agents. It lists the agents you can start a thread with that are not plain models:

Agents from plugins. An agent has its own instructions and tools, and its definition sets its model.
Claude Code, when you have linked a Claude subscription.

A thread that starts on one of these agents keeps it. You cannot switch that thread to another model.

Workspace Settings

A workspace admin can set New Threads Start On, Default Speed for New Threads, and Top Models for every member under Workspace Settings → Model Selector.

Members see a Workspace badge on the settings their workspace chose:

New Threads Start On follows the workspace until a member picks their own, which then wins. Pick Workspace to follow the workspace again. Members can still pick any model for a thread.
Top Models lists the workspace’s models first. Members can add up to 8 models of their own below them.
Puck and Spaces run on the workspace’s models until a member picks their own.

Default Speed for New Threads starts members’ new threads at Standard speed instead of the speed each member used last. Members can still pick a faster speed.

Enterprise workspace admins can also turn off models for their members in Allowed Models. Members cannot pick a model that is turned off. Allowed Models cannot turn off the model new threads start on, so pick another under New Threads Start On first.

Threads That Agents Start

A thread that an agent starts for you uses the model you ask for, e.g., “start it on GPT-6 Astra”. If you name no model, the new thread uses the model of the thread that started it.

Puck and the agent you chat with in a Space are not threads, so the selector does not set their models. Choose a model and effort for each at the bottom of Settings → Model Selector. The threads they start use the model you ask for, or Staff Pick. If Google Painter is enabled for your account, the same section also chooses the Painter image model.

Use a ChatGPT Subscription

When you sign in with ChatGPT and allow token sharing, OpenAI models bill to your ChatGPT subscription instead of Amp credits. The selector shows the ChatGPT icon on the models your subscription serves. Models from other providers still bill to your other connections or to Amp credits.

From the CLI, run amp config model-providers add-chatgpt-subscription (or amp add-chatgpt-sub) to open the same sign-in flow in your browser. Existing legacy ChatGPT connections keep working, and you do not need to reconnect them.

If ChatGPT does not share your plan with Amp, the connection stays off and Model Routing says so. Token sharing is not available on every ChatGPT plan, and ChatGPT Business and Enterprise workspaces can turn it off. In that case, run amp config model-providers add-chatgpt-subscription --legacy on a machine with a browser to sign in to OpenAI directly and add a legacy ChatGPT subscription connection instead.

Costs and Limitations

You can use runners on your own machine for free. When using orbs, you pay for the orbs (from Amp Megawatt/Gigawatt or Amp credits). The orbs are not covered by your ChatGPT plan.

Image generation with Painter still uses Amp credits because OpenAI does not support it through ChatGPT subscriptions. Web search, reading web pages, and voice also use Amp credits.

Use Your Claude Subscription

Claude models that your Claude plan serves run through Claude Code, so they start to respond more slowly than Claude models that Amp serves. See Costs and Limitations.

When you link a Claude subscription in Settings → Model Routing as Claude Subscription, Amp uses your Claude plan for the Claude models you pick in the selector. The selector shows the Claude icon on those models. The Task and Read Thread subagents run the same model as their thread, so they use your plan too. Puck runs Claude Opus 5.5 while your Claude subscription is linked, so it also uses your plan. Orbs and runners use the linked subscription.

A new Claude subscription goes above your API keys in Model Routing, so it serves Claude models before any of them. Drag another connection above it to use that connection for Claude models instead. If your workspace does not allow personal Model Routing, your Claude subscription serves nothing.

Claude Code serves these requests from your Claude plan. It runs through the Claude Agent SDK on Amp’s servers, so nothing needs to be installed in your orb or on your runner. For an overview, read Use Your Claude Subscription in Amp.

Use the Claude Code Agent

The Claude Code agent is another way to use your Claude subscription. It runs Claude Code with Amp’s tools, and it requires a linked Claude subscription. To start a thread with it, open Other Models in the selector and pick Claude Code under Your Agents.

For threads you start from the CLI in execute mode (-x) or orb execute mode (-ox), set the model and effort with --mode and --features:

amp -x "Review this repository" --mode external-agent:claude-code \
  --features model:claude-opus-5-5

amp -ox "Review this repository" --mode external-agent:claude-code \
  --features model:mythos,effort:high

You can pass any model ID supported by Claude Code and available on your plan.

Costs and Limitations

Claude Code serves requests differently from Amp’s direct Claude access:

Each response takes longer to start.
Claude Code adds its own attribution and environment details to the prompt.

These things are not covered by your Claude plan:

Models from other providers, e.g., GPT-6 Astra, use your other connections or Amp credits.
Painter (which generates images) uses Amp credits. To avoid it, add Do not use the painter tool. to AGENTS.md (or CLAUDE.md if present).
You can use runners on your own machine for free. When using orbs, you pay for the orbs (from Amp Megawatt/Gigawatt or Amp credits).

If You Used the Dial

The model selector replaces the Dial and its four modes, low, medium, high, and ultra.

New threads start on Staff Pick, whichever mode you used last.
Existing threads keep working. A thread you started in high continues on the model it ran.
Dial settings do not carry over. If you chose the model behind a mode, pick that model in the selector. If you put your own agents on the Dial, you find them under Your Agents.

If you never changed the Dial’s settings, these are the models your modes ran:

low ran GLM-5.3-Flash.
medium ran Claude Opus 5.5.
high ran GPT-6 Astra.
ultra ran Claude Fable 5.2, which is under Other Models.

Read Bye Bye, Dial for why we made the change.