Esta página aún no está traducida; se muestra la versión en inglés.

Beta Próximamente. Esta función aún está en desarrollo; los nombres y detalles pueden cambiar antes del lanzamiento.

Cost Forecast & Model Router

Know what a request will cost before you send it, and stop paying top-model prices for simple steps. The cost forecast shows an estimate in the composer as you type. The model router sends simple steps of a run, such as reading files or writing a commit message, to a cheaper model, and keeps the hard parts on the model you picked.

The router never routes to Chinese-origin models (DeepSeek, Qwen, Kimi / Moonshot, GLM / Zhipu, MiniMax, Baichuan, Yi and others). This is a fixed rule, not a setting.

Turn it on

  1. Open Settings › Tools › Game changers.
  2. Turn on Cost Forecast.
  3. Turn on Model Router if you also want cheaper models for simple steps. It is off by default.
SettingDefaultEffect
ingot.agent.costForecast.enabledoffshow the estimate in the composer
ingot.agent.costForecast.warnAbove1.00estimated cost in USD that shows a warning before sending
ingot.agent.router.enabledoffsend simple steps to a cheaper model

Cost forecast

As you type, a chip next to the model in the composer shows the estimate for the request:

≈ $0.18 · 42k tokens

The estimate is the expected tokens (your prompt, attached files, rules, chat history and a typical reply) multiplied by the price of the selected model. Hover the chip to see the breakdown. Local models show $0.00. With your own API key, the estimate uses the provider's list price.

Over the threshold

When the estimate is above ingot.agent.costForecast.warnAbove, the chip turns amber and Send asks first:

This request is estimated at $2.40 (310k tokens), above your $1.00 limit.
ChoiceResult
Send anywaythe request is sent as is
Use a cheaper modelpick another model; the estimate updates
Cancelback to the composer, nothing is sent

The forecast is an estimate. The real cost of a run depends on how many steps the agent takes and is shown in Models & limits usage as usual.

Model router

With the router on, every step of a run is sorted before it is sent:

StepModel
reading and searching files, summarising tool output, short edits, commit messagesa cheaper model
planning, multi-file edits, debugging, anything the agent marks as hardthe model you picked
  1. Pick your model in the composer as usual.
  2. Send the request.
  3. Steps that went to a cheaper model get a note on their tool card, with the model that ran them:
Routed to cheaper model
  1. Click the note and choose Rerun on my model if a routed step wasn't good enough.

The router only uses models you already have access to, through Ingot or your own keys, and it never picks a Chinese-origin model, even if one is enabled in your model list.

The model you accept becomes a habit

Ingot keeps track of which model you keep for each kind of task, and which routed steps you rerun on your own model. Once a pattern is clear, the router follows it: for example, it stops routing test fixes if you always rerun them. These preferences are saved with your habits in .ingot/habits.json (git-ignored, never leaves your machine):

Saved to Habits

Silver notes

NoteWhen
Routed to cheaper modela simple step ran on a cheaper model
Saved to Habitsyour model choice for a kind of task was learned