[Bug] OpenRouter provider fails with certain models - returns OPENAI_ERROR

[ankushchhabradelta4infotech-ai]

Description

The OpenRouter provider fails with certain models (GPT-4o, GPT-4o Mini, Grok, Gemini 2.0 Flash, etc.) returning error code OPENAI_ERROR, even though the provider being used is OpenRouter, not OpenAI directly. Other models like Claude and Gemini 2.5 Pro work fine. The openrouter/auto model also works.

Steps to Reproduce

1. Configure the OpenRouter provider with a valid API key
2. Select a model like gpt-4o, gpt-4o-mini, grok-2, or gemini-2.0-flash-001 from the model picker
3. Send any message

Expected Behavior

The request should be routed through OpenRouter to the selected model and return a valid response.

Actual Behavior

The request fails with errors like:

{"type":"error","message":"404 No endpoints found for x-ai/grok-2.","code":"OPENAI_ERROR"}

{"type":"error","message":"400 Provider returned error","code":"OPENAI_ERROR"}

Working models: Claude 3.5 Sonnet, Gemini 2.5 Pro, openrouter/auto

Failing models: GPT-4o, GPT-4o Mini, Grok-2, Gemini 2.0 Flash 001

Root Cause Analysis

After reading the source code, I found two issues:

Issue 1: Error code says OPENAI_ERROR because the provider internally uses the OpenAI SDK

In packages/llm-sdk/src/providers/openrouter/provider.ts, the OpenRouter provider creates an OpenAI client under the hood:

const { default: OpenAI } = await import("openai");
client = new OpenAI({
  apiKey,
  baseURL, // pointed at openrouter.ai
  defaultHeaders: headers,
});

All errors from OpenRouter get wrapped in OpenAI's error format, which is why the error code shows OPENAI_ERROR instead of something OpenRouter-specific.

Issue 2: Hardcoded OPENROUTER_MODELS map has outdated model IDs

The OPENROUTER_MODELS map in both index.ts and provider.ts contains a hardcoded list of only ~20 models. OpenRouter supports 500+ models. Models like Grok are completely missing from this map, and several existing entries have outdated IDs.

Model in SDK	Current OpenRouter ID
google/gemini-pro-1.5	google/gemini-1.5-pro (renamed)
google/gemini-flash-1.5	google/gemini-1.5-flash (renamed)
google/gemini-2.0-flash-exp	google/gemini-2.0-flash-001 (graduated from exp)
Missing entirely	x-ai/grok-2, x-ai/grok-3
Missing entirely	google/gemini-2.5-pro

Issue 3: supportedModels only lists hardcoded models

In index.ts, the provider exposes only the hardcoded keys as supported:

return createCallableProvider(providerFn, {
  name: "openrouter",
  supportedModels: Object.keys(OPENROUTER_MODELS), // Only ~20 models
  getCapabilities,
});

If the UI validates against this list, any model not hardcoded will fail.

Suggested Fix

Replace the hardcoded OPENROUTER_MODELS map with a dynamic fetch from OpenRouter's models API:

// Fetch models dynamically from OpenRouter
async function fetchOpenRouterModels(
  apiKey: string
): Promise<Record<string, OpenRouterModelConfig>> {
  const response = await fetch("https://openrouter.ai/api/v1/models", {
    headers: { Authorization: `Bearer ${apiKey}` },
  });
  const data = await response.json();

  const models: Record<string, OpenRouterModelConfig> = {};
  for (const model of data.data) {
    models[model.id] = {
      vision: model.architecture?.modality?.includes("image") ?? false,
      tools: model.supported_parameters?.includes("tools") ?? true,
      jsonMode:
        model.supported_parameters?.includes("response_format") ?? true,
      maxTokens: model.context_length ?? 128000,
    };
  }
  return models;
}

This way the SDK always stays in sync with what OpenRouter actually supports, and no model ID goes stale.

As a short-term fix, the hardcoded model IDs should at least be updated to match current OpenRouter IDs and include missing popular models (Grok, Gemini 2.5, etc.).

Also the error codes from the internal OpenAI client should be caught and re-mapped to something like OPENROUTER_ERROR so users are not confused.

Code Example

import { openrouter } from "@yourgpt/llm-sdk";
import { streamText } from "@yourgpt/llm-sdk";

// This works
const result1 = await streamText({
  model: openrouter("anthropic/claude-3.5-sonnet"),
  prompt: "Hello!",
});

// This fails with OPENAI_ERROR
const result2 = await streamText({
  model: openrouter("x-ai/grok-2"),
  prompt: "Hello!",
});

Environment

• SDK Version: latest
• Framework: Next.js
• Node Version: 18+

Screenshots / Logs

{"type":"error","message":"404 No endpoints found for x-ai/grok-2.","code":"OPENAI_ERROR"}
{"type":"error","message":"400 Provider returned error","code":"OPENAI_ERROR"}

[Sahil5963]

We've addressed this in beta with live OpenRouter model fetching - no more hardcoded model IDs. You can try it out by installing:

npm install @yourgpt/copilot-sdk@beta @yourgpt/llm-sdk@beta

Let us know if this resolves the issue for you.

[Sahil5963]

🐛 This bug has been fixed and released in v2.1.0. Thanks for reporting!