Automatic model selection

Automatic model selection adds AUTO to the Agent model picker. Instead of choosing a model for every message, you assign a model to each of three levels once. Rusty rates each request and runs it on the model for that level, so quick questions stay fast and cheap while hard problems get a stronger model.

How it works

1. You choose AUTO in the Agent model picker.
2. A decision model, called JEV in the app, rates the request as Light, Standard or Heavy. It never sees your model list, so your mapping alone decides what runs.
3. The model you assigned to that level runs the request.

When the decision model is unsure and leans toward a higher level, AUTO steps up one level. Under-powering a request costs more than over-paying for it.

Level	Typical requests
Light	Conversation, factual questions, explaining code, and small single-file edits.
Standard	Features and bug fixes across a few files, tests, reviews, and debugging with a clear error.
Heavy	Architecture, large refactors, subtle concurrency or performance bugs, and long multi-step tasks.

Workflows: a model for each step

A workflow's steps differ a lot in difficulty. Research and Verify read and report, while Architect and Debug reason hard, so rating the whole message once would waste a strong model or starve a hard step. When AUTO is selected and the chat follows a workflow, the choice is made per step instead:

1. The run starts on your Standard model. Rusty does not rate the message up front.
2. Just before each agent step's first model turn, JEV rates that step as Light, Standard or Heavy. It sees the step's instructions, your request and the hand-offs from earlier steps, and the step runs on the model for that level.
3. Every turn of the step then uses that model. The next step is rated on its own.

The chat shows each choice, for example ↳ AUTO · Light task · your-model (90% confidence). As with single requests, an unsure rating steps up one level.

Some limits to know:

• A step can only switch to a model on the same provider the run started on. If a level's model is on another provider, or on a managed-login provider, that level is skipped and the nearest level on the same provider is used. Assign all three levels on one provider to get the full range.
• If the decision service is slow (over about 12 seconds) or fails, the step simply keeps your Standard model. Choosing never stops a step.
• Each rating costs one Decisions API request.

Set it up

Requirements: an OpenRouter API key, and a JEV Decisions API model available in your OpenRouter model book.

1. Connect OpenRouter in LLM Integrations with your API key. See Model providers.
2. Refresh the OpenRouter model catalog so the decision models appear.
3. Open Settings → Intelligence → Intelligent model selection.
4. Under JEV decision model, choose a version. All compatible versions from your authenticated OpenRouter model book are listed.
5. Assign a model to each of the three levels. You can use any model from any connected provider, reasoning-effort variants included.
6. Turn on Enabled.

The panel shows a status line. Ready means AUTO is available. If it is not, the message tells you what is missing: connect OpenRouter, refresh the model book, or choose a model for a specific level.

Tip

Start with a fast model for Light, a balanced one for Standard and your strongest model for Heavy. Adjust after you see how the decisions play out.

See what it decided

Every AUTO decision is recorded in Tool Execution Observability, including the request, the probabilities for each level, the confidence, the model chosen and what the decision cost. In a workflow, each step's decision is recorded separately as AUTO model selection · step. See Observability & metrics.

Related: choosing the workflow

The same decision model can also choose the workflow for a message, and hand a running workflow over to another. That is separate from choosing a model, and the two combine: Auto picks the workflow, then AUTO picks a model for each of its steps. See Let Rusty choose the workflow. Each routing decision is recorded in Tool Execution Observability as AUTO flow selection or AUTO flow switching.

Related experimental features

The same decision model powers three experimental features, also under Settings → Intelligence:

• JEV decisions (experimental). Agents lay out options and JEV picks one. In AUTO chats it can also move a chat up to a stronger model when the remaining work needs it. A setting controls the confidence at which Rusty acts on JEV's pick.
• JEV review of risky actions (experimental). Before an agent overwrites a file destructively or runs a destructive command, JEV reviews the step. It can let it run, send the agent back to revise, or ask you.
• JEV decision review (experimental). A shadow mode that scores every tool call any model makes as Proceed, Revise or Stop. The call always runs and the verdict is only recorded in Tool Execution Observability. Each scored call is one Decisions API request.

All three need OpenRouter connected with a decision model chosen, and all may change between releases.