Missleading documentation or incorrect behaviour about small_model

Description

Looking at the documentation for OpenCode at https://opencode.ai/docs/config/:

The small_model option configures a separate model for lightweight tasks like title generation. By default, OpenCode tries to use a cheaper model if one is available from your provider, otherwise it falls back to your main model.

However

opencode/packages/opencode/src/provider/provider.ts

Line 1109 in 253b7ea

export async function getSmallModel(providerID: string) {

The code shows that OpenCode first attempts to use its own opencode provider for the small model. Only if the opencode provider is unavailable does it fall back to the user's main configured model.

This discrepancy can lead to unexpected behavior. For example, when using deepseek-reasoner via the DeepSeek provider as a main model, OpenCode silently used opencode nano-5-gpt as the small model. This means session data was sent to OpenCode's servers without explicit notice, raising potential privacy concerns.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

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

[BUG] Default small_model charges users when free alternative exists on Copilot #6802: [BUG] Default small_model charges users when free alternative exists on Copilot - addresses the same problem of small_model silently using OpenCode's provider without explicit user knowledge
Github Copilot burns premium quota for non-premium models #6816: Github Copilot burns premium quota for non-premium models - related issue about unexpected model selection causing unwanted charges

Feel free to ignore if your specific case differs from these existing reports.

This means session data was sent to OpenCode's servers without explicit notice, raising potential privacy concerns.

If you don't want opencode provider to be active, you can disable it in your config:

{
  "$schema": "https://opencode.ai/config.json",
  "disabled_providers": ["opencode"]
}

Or it's also possible to configure a whitelist-only using enabled_providers.

What I agree is done incorrectly in the linked functions is this:

    if (providerID.startsWith("github-copilot")) {
        // prioritize free models for github copilot
        priority = ["gpt-5-mini", "claude-haiku-4.5", ...priority]
    }

claude-haiku-4.5 has a cost of 0.33 premium requests per request, it's absolutely not free; if more models are needed than just gpt-5-mini, then gpt-4.1 is another candidate

Yes, my point is not about how to configure it, but that the stated default behavior in the documentation contradicts the actual code.

More importantly, the actual default behavior—silently sending data to a third-party service (opencode) when the user has configured a different primary provider—constitutes a privacy issue and should be clearly communicated.

I don't see any silent sending of data to 3rd-party services. I only see that if you already use opencode provider, you get the most appropriate small model for it. If I use github-copilot, I'll get the most appropriate model for github-copilot, the check for opencode is just earlier but not relevant (EDIT: nor applied) to my use in that case.

I don't see any silent sending of data to 3rd-party services.

I launched opencode CLI, chose DeepSeek as the provider, and started using it without any additional configuration. I expected that my session and my data would be sent only to DeepSeek. However, for some reason, my session data is being sent to opencode as well.

That's what I mean by "silent sending data to 3rd-party services"

fix: remove opencode fallback in getSmallModel

This opencode behaviour is a problem for me too, and I found this bug while looking for how the session titles were constructed.

I want to use opencode in two modes:

vibe coding things which are 99% public anyway
editing my personal notes which I want to stay private

In order to do this I have:

coding models (using various providers, including opencode)
local models (using locally hosted models, various servers)

When editing my notes I'm expecting slower responses, and this is a tradeoff I've made for privacy.

How the session titles were generating so fast was suspicious.

It seems that at the moment, if I have "other providers" than my local models, some parts of the sessions with the local models may leak to those providers.

Is there a way for me to configure opencode that I can meet both my use cases, or do I need to maintain multiple config files?

This behaviour is documented, albeit under the self-hosted GitLab provider:

opencode/packages/web/src/content/docs/providers.mdx

Lines 687 to 701 in 712d2b7

    
           :::note[compliance note] 
        
           OpenCode uses a small model for some AI tasks like generating the session title. 
        
           It is configured to use gpt-5-nano by default, hosted by Zen. To lock OpenCode 
        
           to only use your own GitLab-hosted instance, add the following to your 
        
           `opencode.json` file. It is also recommended to disable session sharing. 
        
           ```json 
        
           { 
        
             "$schema": "https://opencode.ai/config.json", 
        
             "small_model": "gitlab/duo-chat-haiku-4-5", 
        
             "share": "disabled" 
        
           } 
        
           ``` 
        
           :::

The small_model config docs should probably be updated with something along the same lines.

It seems like the fallback to gpt-5-nano from the opencode provider has recently been removed, see 7d7837e. With that change, the small model should always be selected from the current provider. So this issue can be closed, I guess?

Woe. that is indeed conceraning. Furthermore it is sent to OpenAI wgich stores data for 30 days according to domcumetation of OpenCode service. And code isn't actually searching for smaller model present, it search against predefined list.

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.

Navigation Menu

Repository navigation

Missleading documentation or incorrect behaviour about small_model #8609

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	:::note[compliance note]
	OpenCode uses a small model for some AI tasks like generating the session title.
	It is configured to use gpt-5-nano by default, hosted by Zen. To lock OpenCode
	to only use your own GitLab-hosted instance, add the following to your
	`opencode.json` file. It is also recommended to disable session sharing.

	```json
	{
	"$schema": "https://opencode.ai/config.json",
	"small_model": "gitlab/duo-chat-haiku-4-5",
	"share": "disabled"
	}
	```

	:::

Missleading documentation or incorrect behaviour about small_model #8609

Description

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Jan 15, 2026

Nindaleth commented on Jan 15, 2026

podliy16 commented on Jan 15, 2026

Nindaleth commented on Jan 15, 2026

podliy16 commented on Jan 15, 2026

davedean commented on Jan 20, 2026

tlvince commented on Jan 27, 2026

ghost91- commented on Mar 11, 2026

b-m-tr commented on May 6, 2026

github-actions commented on Jul 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions