Amp Replaces Four Dial Modes With Direct Model Selection in the Thread
Sourcegraph's coding agent Amp has removed the Dial switch with its four difficulty modes and returned to a plain model picker. The reason is billing: after Amp added support for users' own subscriptions and API keys, the medium mode could charge Claude, ChatGPT, or Amp credits, and it became hard to tell which. Each model in the list now shows who pays, and you can change models in the middle of a thread. New threads start on Staff Pick by default, which is currently Claude Opus 5.5.
AI-processed from Amp News (Sourcegraph); edited by Hamidun News
The team behind Amp, Sourcegraph's coding agent, has admitted it changed its mind. The Dial switch with the low, medium, high, and ultra modes is going away, and a regular model list is taking its place. For people who work with agents every day, it's a clear example of how a neat abstraction breaks down when money is involved.
Why Dial stopped working
The idea behind Dial was simple. Users think about how hard a task is, not about model names and versions. In January, Amp explained why it didn't offer model choice at all. In July, it released Dial. According to the team, it worked: when the model behind a mode was swapped, nobody complained.
Then Amp added the ability to connect your own subscriptions and API keys. With those, Amp becomes free to use. The selected model then determined not only who generates tokens, but also who pays for them. The medium mode could charge your Claude plan, your ChatGPT plan, or your Amp credits. It became very hard to tell who was paying for which model at any given moment. On top of that, the team had to keep adding settings around Dial so it would work across every combination.
So the models stopped being hidden. In the new selector, each row has a badge: Claude, ChatGPT, or Amp. It shows whose account will be billed. A card next to the list describes the model and states, for example, that it runs on your personal Claude subscription.
What users get
The choice isn't entirely bare. The team stresses that the agent is more than the model. For each model in the list, Amp prepares its own system prompt and toolset, and these change along with the model.
The main new feature is switching models in the middle of a thread. Previously, a thread stayed on the mode and model it started with, because a model could get confused by continuing a conversation begun by another one. Now, according to Amp, models are good enough to handle that handoff. After a switch, the history gets an entry such as "Switched to GPT-6.1 Sol · High," and the new model answers the next message. For each thread you can choose not only the model but also the effort level.
There is a cost. The prompt cache belongs to a single model, so after a switch the new model reads the entire thread without cache, and the first message costs more. Threads in which the agent runs from a plugin stay with their own agent.
For those who don't want to choose, there is Staff Pick, the model the team would pick itself. Right now that is Claude Opus 5.5. Amp keeps comparing models on benchmarks and promises to change the pick when something better appears. New threads start on Staff Pick by default, unless you previously set a default mode in Dial. In the Model Selector settings, you can start new threads on the last model you used or on one fixed model.
What happens to your settings
- New threads start on Staff Pick, regardless of which mode you used last.
- Existing threads keep working: a thread started in high stays on the model it was running.
- Dial settings don't carry over. If you changed the model behind a mode in Tune Modes, pick it again in the selector. Custom agents from Build Dial now live under Other Models.
If you changed nothing, the modes mapped to these models: low was GLM-5.3-Flash, medium was Claude Opus 5.5, high was GPT-6 Astra, and ultra was Claude Fable 5.2. The first three are moved to the top of the list, while Claude Fable 5.2 sits under Other Models.
What follows from this
The lesson for anyone building their own tools around LLMs: once a product has several payment sources, hiding the model behind an abstraction like "task difficulty" stops being convenient. Users need to see who pays and what exactly is running. The second practical point concerns cost: switching models in the middle of a long thread gets more expensive because the cache is lost, so change models deliberately rather than out of habit. For long sessions, it makes more sense to stick with one model and switch only when the task actually changes.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
Key Questions Answered
Why Dial stopped working?
The idea behind Dial was simple. Users think about how hard a task is, not about model names and versions. In January, Amp explained why it didn't offer model choice at all. In July, it released Dial. According to the team, it worked: when the model behind a mode was swapped, nobody complained. Then Amp added the ability to connect your own subscriptions and API keys. With those, Amp becomes free to use. The selected model then determined not only who generates tokens, but also who pays for them.
What users get?
The choice isn't entirely bare. The team stresses that the agent is more than the model. For each model in the list, Amp prepares its own system prompt and toolset, and these change along with the model. The main new feature is switching models in the middle of a thread. Previously, a thread stayed on the mode and model it started with, because a model could get confused by continuing a conversation begun by another one. Now, according to Amp, models are good enough to handle that handoff.
What happens to your settings?
New threads start on Staff Pick, regardless of which mode you used last. Existing threads keep working: a thread started in high stays on the model it was running. Dial settings don't carry over. If you changed the model behind a mode in Tune Modes, pick it again in the selector. Custom agents from Build Dial now live under Other Models. If you changed nothing, the modes mapped to these models: low was GLM-5.3-Flash, medium was Claude Opus 5.5, high was GPT-6 Astra, and ultra was Claude Fable 5.2.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.