Openrouter unwated requests to Claude Haiku 4.5.

Description

I have been using OpenCode with free models on openrouter for example Grok Code fast 1 and currently Grok 4.1 Fast but on openrouter dashboard i can some calls to Claude Haiku 4.5 through the opencode app that i never did, they are very small with just 10-300 tokens but slowly slowly the end up building up to different cents.

OpenCode version

1.0.80

Steps to reproduce

Connect OpenRouter
Message with free models like Grok 4.1 fast or grok code fast 1

Screenshot and/or share link

Operating System

Fedora 42

Terminal

Alacritty + Fish

@Yyeger

Hey @Yyeger Haiku must be used for things like Session Title Generation, I think adding this to you config you can switch it to another model a free one maybe.

{
  "$schema": "https://opencode.ai/config.json",
 ...
 "small_model": "x-ai/grok-4.1-fast:free"
}

^^ yup

there are free openrouter models too i think, u can also disable this by doing

small_model: “”

or you can use free opencode models:

opencode/big-pickle

Bump. This is not a reasonable default. If I can and do select a (main) model, I expect that model to be used.

Three issue's with the current approach

is this actually necessary? Do I need title naming and whatnot?
model selection should include small model selection and never default to a model I explicitly do not want to use
on the ui I can clearly see my model selection for the main model but cannot see anywhere the small model selector. Unless I trawl though the logs from open router I wouldn't even know opencode is hammering a model I didn't select

@adam-ah

I agree with @adam-ah. It's money right out of my pockets. It took me some time to realize what was going on, and it’s definitely not good. Even when you use a free model, you still get charged by default. Seriously? Fortunately, the solution is straightforward: you can either assign the small_model to match the model you're already using or choose another free option, but this should really be the default behavior. For those interested, I’ve developed an AI CLI wrapper where I addressed this for OpenRouter, it automatically sets the small_model to match your selected one. It’s called Omni-CLI.

there are free openrouter models too i think, u can also disable this by doing

small_model: “”

I just tested, keeping small_model empty still defaults to anthropic/claude-haiku-4.5

only something like small_model: "something/invalid" works.

Also, this feature sends every message in your session (one by one, as you send the message) to this model, which might be bad for privacy for some people, if they don't want to keylog their data to this particular provider

Not to mention the cost

@TheHamkerCat

Thanks @TheHamkerCat for the tip!
I think small_model should be disabled by default, the current behavior looks unreasonable.

(fix): Disabling small_models due to anomalyco/opencode#4579

I also think unexpected call to small_model should be disabled by default. The default should be the model selected if its really necessary

I was also just recently confused by this same behavior, phantom charges on my openrouter account due to opencode using a model I NEVER picked and don't even have enabled, to do something I didn't expect it to do.

Not only is this a potential issue with opencode, it nearly sent me to openrouter customer support to claim that I was not in fact using haiku for anything, and claim that they may have an issue on their end.

ie, this is such a problem that it may unduly burden providers support desks because the user will think they're getting charged incorrectly for something that they're not using.

small model should have an on/off switch in the UI
the model to use should have a picker, or should default to the model you're using for the request, if unset
the UI should make it clearer that it's sending requests for session title generation.
It would be nice if there was some option for this to be done locally using some sort of tiny model

Absolutely no need whatsoever for the end user to specify that they want to use a free model. OpenCode could just default to that, and not use a paid model like Haiku. It looks like it's done on purpose too. What's even funny is that my OpenRouter key is set to have a preset with just Deepseek v4 Pro (using only the DeepSeek official instance), OpenCode completely disregarded that and used Haiku. Very sketchy.

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

Openrouter unwated requests to Claude Haiku 4.5. #4579

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Three issue's with the current approach

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Openrouter unwated requests to Claude Haiku 4.5. #4579

Description

Description

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

OpeOginni commented on Nov 21, 2025

rekram1-node commented on Nov 21, 2025

rekram1-node commented on Nov 21, 2025

adam-ah commented on Jan 17, 2026

Three issue's with the current approach

mabelisle commented on Jan 21, 2026

TheHamkerCat commented on Jan 31, 2026

txchen commented on Feb 6, 2026

shahidcodes commented on Apr 3, 2026

drobin04 commented on May 16, 2026

classy-giraffe commented on Jun 8, 2026

github-actions commented on Aug 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions