Proposal: Reasoning option / budget

[EFLKumo]

Considering the two major approaches currently used to configure reasoning options for reasoning models—setting reasoning efforts such as low, medium, or high, and setting reasoning budgets—I believe it's necessary to maintain which approach each model supports, along with the enum of reasoning efforts or budget limits. Either a model's reasoning behavior could be turned off or not might also be taken into consideration.

Thank you for your tremendous contributions to this great database.

[rubnogueira]

I was going to open an issue regarding the same subject, but I just found this here.

I'm facing inconsistencies in my implementation of reasoning effort control usage. Furthermore, I use this list as a source for my models, and there are many approaches.

Types:

• Some always use reasoning.
• Some never use reasoning.
• Some use reasoning, but it's possible to turn it off.
• Some only use reasoning and allow adjusting the effort, or you can decide if you want to dynamically.
• Some allow reasoning or not and allow adjusting the effort, or you can decide if you want to dynamically.

Models:

• gemini-2.5-pro: can't turn reasoning off; allow a budget of 128 to 32768 while -1 is dynamic thinking. Default dynamic.
• gemini-2.5-flash: budget 0-24576, 0 thinking off, -1 dynamic thinking. Default dynamic.
• gpt-5: can select between low/medium/high. Default is medium.
• gpt-5-pro: only supports high
• gpt-4.1: no reasoning at all
• grok-3-mini-fast: support low/high values
• grok-4-fast: has reasoning, but no control
• grok-4-fast-non-reasoning: no reasoning at all

I feel that we need to find a schema or a way to represent all these constraints. @rekram1-node sorry to tag you directly, but I saw your recent merge requests. Do you think we can think about a solution for this? Thanks!

References that I used:

https://ai.google.dev/gemini-api/docs/thinking#set-budget
https://platform.openai.com/docs/api-reference/chat/create#chat-create-reasoning_effort
https://ai-sdk.dev/providers/ai-sdk-providers/xai#provider-options
https://docs.x.ai/docs/guides/reasoning#control-how-hard-the-model-thinks

[rekram1-node]

@rubnogueira no problem you are always free to tag me!

Yeah there is definitely some confusion on how to configure reasoning and tbh I think it'd be better if we could do most of it in the tui only so a user doesnt have to concern themselves with the insides of how it works.

Right now it is super non obvious and not documented but to change reasoning settings I just lookup the provider in vercel ai sdk docs

Example: openai

Then I read how to configure reasoning, example:

reasoningEffort 'minimal' | 'low' | 'medium' | 'high' Reasoning effort for reasoning models. Defaults to medium. If you use providerOptions to set the reasoningEffort option, this model setting will be ignored.

reasoningSummary 'auto' | 'detailed' Controls whether the model returns its reasoning process. Set to 'auto' for a condensed summary, 'detailed' for more comprehensive reasoning. Defaults to undefined (no reasoning summaries). When enabled, reasoning summaries appear in the stream as events with type 'reasoning' and in non-streaming responses within the reasoning field.

Then I put it in my config:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "openai": {
      "models": {
        "gpt-5": {
          "options": {
            "reasoningEffort": "low",
            "reasoningSummary": "auto",
            "include": ["reasoning.encrypted_content"]
          }
        }
      }
    }
  }
}

I know this sucks and it isn't ideal and we are thinking of ways to make it better