Amp Killed the Dial. You Now Pick a Model per Thread and See Who Bills It. Here Is the Routing Rule I'd Set Up.
Amp, the coding agent, published a post on November 12 called "Bye Bye, Dial." The Dial was its four-mode switch, low, medium, high and ultra, which let you choose by how hard a task was instead of by model name. It's gone. In its place is a model selector where you pick the model and effort for each thread, and every row shows who pays for it: one of your subscriptions, one of your API keys, or Amp credits. For a solo developer paying for Claude and ChatGPT already, that payer label is the most useful thing a coding agent has shipped in a while.
Why the Dial stopped making sense
Amp's own explanation is about billing. Once Amp let you bring your own Claude subscription, ChatGPT subscription or API keys, the same mode could end up billed to different places depending on which model sat behind it. "Medium" didn't tell you whether you were spending your Claude plan's weekly allowance or Amp's credits.
If you never changed your Dial settings, the old modes map to models like this: low was GLM-5.3-Flash, medium was Claude Opus 5.5, high was GPT-6 Astra, and ultra was Claude Fable 5.2. New threads now start on a "Staff Pick," which is Claude Opus 5.5 today, and you can change the default to your last choice or a fixed model in settings. Existing threads from the Dial era keep the model they had.
Who pays for what
The billing rules are worth reading closely, because they're where the money goes.
If you link a Claude subscription, Amp uses your Claude plan for the Claude models you pick, and its model routing view shows how much of your five-hour and weekly limits is left. If you sign in with ChatGPT and allow token sharing, OpenAI models bill to your ChatGPT subscription instead of Amp credits. Amp notes token sharing isn't available on every ChatGPT plan.
Some things still use Amp credits even when a subscription covers the model: image generation, web search, reading web pages and voice. Any model that none of your own connections serve also bills to credits. So "I linked my subscriptions, this is free now" isn't quite right. The model tokens move to plans you already pay for, and the tool calls around them still cost something.
The hidden cost of switching mid-thread
You can switch model or effort in the middle of a thread, and the transcript records it. Amp is upfront about the catch: prompt caches belong to one model, so after a switch the new model reads the whole thread uncached, and the first message after the switch costs more than usual.
On a long thread that's not a small number. Cached input is usually billed at a fraction of the normal input price, so a 150,000-token thread re-read at full price can cost more than the next ten cached messages combined. If that lands on your Claude plan, it's a chunk of your five-hour window. If it lands on credits, it's money.
There are exceptions to switching, too. Plugin agents and Claude Code threads keep their agent and can't switch model, and in the CLI only the thread owner can switch.
The routing rule I'd set up
Here's the rule I'd use as a solo developer with both subscriptions linked.
Pick the model when you open the thread, not halfway through. Decide up front whether the task is a quick edit, a normal feature, or a hard debugging session, and open the thread on the model you'd want at the end. Switching is for when the first model is clearly failing, not for saving a few cents on the easy part.
Put routine work on the plan with the most headroom. Amp shows your remaining Claude limits. If you're near the end of the weekly window on Thursday, open new routine threads on an OpenAI model that bills to ChatGPT, and keep the Claude allowance for the work where you think Claude does better.
Use the cheap model only for tasks that are small and self-contained: renaming, writing a test for an existing function, a one-file refactor. Those are the threads where you won't be tempted to switch later.
Start a fresh thread instead of switching when the task changes. A new thread with a short summary of where you are is cheaper than dragging a long history into a model that has to read all of it uncached.
Where I could be wrong
I'm assuming most solo developers care more about which bill gets hit than about squeezing out the best model for every message. If you're on API keys with a generous budget, the Dial's simplicity might have been worth more than the billing clarity, and the selector is just more decisions.
I'm also guessing at the size of the switching penalty from typical cache pricing, not from Amp's numbers. Amp doesn't publish per-model token prices on its pricing page, so I can't tell you exactly what a switch costs in credits. Run one long thread, switch once, and look at the usage view before you take my rule as settled.
What I'd actually do
If you use Amp, set "New Threads Start On" to the model you use most, link both subscriptions if your ChatGPT plan supports token sharing, and spend a week opening every thread on a deliberate choice. At the end of the week, check where your usage actually went. If you use another coding agent, ask whether it can show you who pays per model. In late 2026 that label is the difference between knowing your AI bill and guessing at it.
Author
Lukas
@lukcombinator