Automatic Model Routing in OpenCode Mobile

Auto model routing is opt-in. When it is enabled, OpenCode Mobile classifies each new prompt as Simple, Standard, or Complex and sends it with the model mapped to that tier. It is useful when you want routine checks on a fast model without manually switching models for every turn.

Routing never guesses a tier with a heuristic. If the classifier is unavailable, the send uses your normal default model.

Before you enable it

• Connect to the server and confirm its model catalog is loaded.
• Configure at least one provider and enable the models you want routing to consider.
• Decide which classifier to use: on-device or OpenCode server.

Choose a classifier

Open Settings → AI & providers and find the auto-routing controls.

On-device classifier

The private option. The app downloads a small instruction model and classifies the prompt locally.

• The pack is about 491 MB and is downloaded only after you choose it. It is never bundled with the app.
• Classification works on both OpenCode 1.x and 2.x because it does not call the server.
• The model is released after five minutes idle or when the app backgrounds.
• If the pack is missing, removed, or fails to load, routing is skipped for that send.

OpenCode server classifier

The server option uses the connected OpenCode 2 server’s experimental text-generation endpoint.

• The classifier prompt is sent to the selected classifier model on your server before the main prompt.
• The endpoint is probed before the option is enabled. A missing endpoint, authentication failure, or unsupported server leaves the option unavailable.
• A send-time classifier failure falls back to your default model; it does not switch to the on-device classifier or invent a tier.
• OpenCode 1.x does not offer this option.

Map the tiers

Each tier can stay Auto or point at an explicit model:

Tier	Automatic suggestion
Simple	The cheapest enabled model that supports tools
Standard	Your current default model, falling back to the simple suggestion
Complex	The most capable enabled model: reasoning first, then context size, then price

You can also choose an optional thinking level or model variant for each tier. A variant applies only when the routed model exposes it; otherwise routing continues with the model default.

Automatic suggestions only consider enabled, configured, non-deprecated models. If an explicit tier model becomes unavailable, that send skips routing rather than falling back to a different tier.

Protect long chats

Prompt-cache protection is enabled by default. Above roughly 32,000 context tokens, a downgrade to a cheaper simple or standard model is skipped unless the target model’s input price is at most half of the current model’s. Complex upgrades are still allowed.

This avoids losing a warm prompt cache just to save a small amount on a long conversation. Turn it off in the auto-routing settings if you prefer cost savings over cache reuse.

What you see in Chat

• The model picker shows an Auto-routing row when the selected classifier is available.
• Choosing an explicit model turns routing off for that chat. Choose Auto-routing again to restore it.
• Each assistant reply shows the model that actually answered it, including auto-routed replies. This is the record of the routed choice after a later turn switches models.
• Settings reports the classifier state, pack status, tier mapping, and cache guard.

When routing is skipped

Routing is deliberately skipped when:

• auto-routing is off;
• an explicit model is selected;
• the session is not idle;
• the prompt is empty or too long for the classifier;
• no visible models are available;
• the selected classifier is unavailable;
• attachments or the estimated context do not fit the selected model;
• the routed model is already the session model;
• the selected tier is unavailable;
• the long-chat cache guard rejects a downgrade.

A skipped route is not an error. The prompt is sent with the normal model selection.

Privacy

• On-device classification never sends the prompt off the device.
• Server classification sends the classifier prompt to the classifier model on your OpenCode server before the main prompt.
• The app does not log or persist prompts or classifier results for training.

Troubleshooting

• The Auto-routing row is missing: confirm the classifier is available and the model catalog has loaded.
• On-device routing is unavailable: download the pack, or remove and reinstall it if it is corrupt.
• Server routing is unavailable: confirm OpenCode 2, the experimental endpoint, and the classifier model.
• A reply used the default model: routing was skipped or the classifier failed for that send. Check the settings status and try again.
• A routed model is missing: enable it in Providers and models, or map the tier to another model.

Next: Providers and models, Start and monitor a task, or Manage extensions.