[FEATURE]: Specify small_model per provider

[rmagatti]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

The Problem

Essentially, I go between multiple different providers, and for each one of them, the small model that I would like to use for tools like Explore, for example, is potentially different.
For example:

• Anthropic: The small model I would like to be Haiku 4.5.
• OpenAI: If I'm using the OpenAI provider, I want it to be something like GPT-5 mini.
• OpenCode Zen: If I'm using OpenCode Zen, I can then set the small model to some free or open source, cheaper version as well.

Usage and Configurability

To outline how the usage and configurability of this new feature would actually work, as per what we already did here, we just added the ability to define the small model directly on the configured provider.
Here is what that looks like in the config:

{
  "providers": {
    "anthropic": {
      "model": "claude-3-7-sonnet-latest",
      "smallModel": "claude-3-5-haiku-latest"
    },
    "openai": {
      "model": "gpt-4o",
      "smallModel": "gpt-5-mini"
    },
    "opencode-zen": {
      "model": "default-zen-model",
      "smallModel": "meta-llama-3-8b-instruct"
    }
  }
}

How it actually works in practice:

• Instead of having one global small model setting, the configuration lives on the provider level.
• When I switch my active provider, sub-agents and tools automatically pick up the correct, specific small model for that environment.
• This prevents me from having to manually reconfigure the small model every time I switch providers to avoid unnecessary costs or errors.
  So the feature request is essentially the ability to specify a small model per configured provider. Let me know what you think!

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [FEATURE]: allow skills to request for small/medium/large model hint #16082: Proposes allowing skills to specify a preferred model size (small/medium/large) with per-provider model size mappings — closely overlaps with the idea of specifying a smallModel per provider
• [FEATURE]: Dynamic model selection for subagents via Task tool #6651: Proposes dynamic model tier selection for subagents, including per-provider model tier configuration

These cover similar ground to the per-provider smallModel concept you're describing. It may be worth coordinating or building on those discussions.

[ghost91-]

At my company, we would also like this functionality. Our use case is slightly different, but quite similar:

We have a custom provider (OpenAI compatible) and we have a small OpenCode plug-in that automatically detects the available models for that provider and adds them to the configuration.

We would also like to automatically configure the small_model used for session title generation. But only, when the user/employee actually uses our provider. There can be cases where they have other providers configured, too, e.g. for work for customers. In those cases, we do not want users to accidentally use a model from our provider, not even for session title generation.

At the moment, the selection of the small model is fairly rigid:

• use model of the title agent, if available
• use small_model, if specified
• use a „cheap“ model of the current provider (cheap is a hard coded list of model names), if available
• use gpt-5-nano of the opencode provider (if it’s not disabled)
• use the main model

What we need is some way to control with a plug-in which small model is used when our provider is used, without impacting it for other providers.

I see a couple of ways of making that possible:

1. Add small_model as a per-provider option, too. The global small_model would still take precedence over this. This is basically what is described in this issue.
2. Add a way to configure a list of „cheap“ models per provider, that are considered as small models. This is similar to the previous option, but allows to configure multiple „candidates“
3. Add a plug-in hook for the title generation, that allows to select the model. Probably not the best idea because it is quite invasive. Just wanted to mention it for completeness.
4. Add a flag small to the model config options. Then, before searching for a „cheap“ model, we would look for a model of the current provider with this flag set to true.

I prefer option 1, but I'm open to explore the other ideas, too, if there is a preference for any of them.

[ghost91-]

Question: In your config example, you also specify the model per provider. How would that work? In OpenCode, you cannot actually really select a provider, you select a model. So the per provider model doesn't make much sense, to me. Or am I missing anything?