Open router support

[ZuidVolt]

Currently, I have been using the open-router API as a custom endpoint for the open-ai provider, which give access to a larger amount of models through the OpenAI API with different providers.

I was wondering if there would be any chance Open router could be a separate model provider in the future because as of right now only custom specified non-OpenAI models work in the assistant panel, as I get this error when trying to OpenAI model though Open Router I get this error message "Error interacting with language model data did not match any variant of untagged enum ResponseStreamResult" which seems to be a zed side issue.

If this requested out of scope for the current project, please close this discussion, thank you for considering!

[awoimbee]

This is not only affecting OpenAI models, I get the same error using anthropic/claude-sonnet-4.
It systematically happens when the model tries to read a file.

Relates to #24067, #30174, #30188, #16576

Edit:
It happens on anthropic models when using the "OpenAI" provider in the settings. I am now using #29496 where anthropic models work great but other models don't integrate well with the 'agent' stuff

[Go36625090]

I have same error..

[taoeffect]

It happens on anthropic models when using the "OpenAI" provider in the settings

This is happening for me now too with OpenAI models like o4-mini (high).

I've verified that this is not an issue with OpenRouter by manually chatting with the same model in the OpenRouter UI.

In case it's helpful, this is my config:

  "language_models": {
    "openai": {
      "api_url": "https://openrouter.ai/api/v1",
      "available_models": [
        {
          "name": "openai/o4-mini",
          "display_name": "o4-mini (OpenRouter)",
          "max_tokens": 200000
        },
        {
          "name": "openai/o4-mini-high",
          "display_name": "o4-mini-high (OpenRouter)",
          "max_tokens": 200000
        },
        {
          "name": "openai/o3",
          "display_name": "o3 (OpenRouter)",
          "max_tokens": 200000
        },
        {
          "name": "openai/gpt-4.5-preview",
          "display_name": "GPT-4.5 (Preview) (OpenRouter)",
          "max_tokens": 128000
        },
        {
          "name": "google/gemini-2.5-pro-preview",
          "display_name": "Gemini 2.5 Pro Preview (OpenRouter)",
          "max_tokens": 1048576
        },
        {
          "name": "anthropic/claude-sonnet-4",
          "display_name": "Claude Sonnet 4 (OpenRouter)",
          "max_tokens": 200000
        },
        {
          "name": "anthropic/claude-sonnet-4:thinking",
          "display_name": "Claude Sonnet 4 Thinking (OpenRouter)",
          "max_tokens": 200000
        },
        {
          "name": "anthropic/claude-opus-4",
          "display_name": "Claude Opus 4 (OpenRouter)",
          "max_tokens": 200000
        }
      ],
      "version": "1"
    }
  },

[JostCrow]

It is now in the last preview build!

[taoeffect]

Does o4-mini (high) work?

[JostCrow]

I have only tested some free models and they work fine. So my guess is that o4 high would work too

[taoeffect]

So my guess is that o4 high would work too

No, I tested other models and they work, there seems to be something special about it, see issue: #30174 (comment)

[robertoandrade]

I wonder how one could configure it to use "middle-out" compression/transform as per OpenRotuer's docs: https://openrouter.ai/docs/features/message-transforms

[SuperKenVery]

Hi, is it possible to select the model provider from Zed?

[JostCrow]

Last I checked, no. But that was 2 weeks ago

[imumesh18]

Currently not there is a open pr. But if you have a specific model<>provider combination which you want to enforce you can create a key for that provider and it should just work(it's a hack for you to get unblocked)

[SuperKenVery]

Thanks, I found the pr: #34500

However I think a GUI option might be better... maybe I'll implement that someday :D