Your demo didn't fail
because of the model.
It failed because your API key
got rate-limited.

// Auto-pick: no provider/model needed.
// AKM picks the best available group and cascades on failure.

const result = await scheduler.withRetry({
  execute: async ({ apiKey, provider, model, signal }) => {
    // Use the injected values — they change on fallback
    return generateContent({ apiKey, provider, model, prompt, signal })
  }
})

// Targeted: start with a specific route, auto-fallback if it breaks

const result = await scheduler.withRetry({
  provider: "openrouter",
  model: "google/gemma-4-26b-a4b-it:free",

  // Optional explicit fallback list
  fallbacks: [
    { provider: "google", model: "gemini-2.5-flash" },
  ],

  execute: async ({ apiKey, provider, model }) => {
    return callSDK({ apiKey, provider, model, prompt })
  }
})

import { ChatOpenAI } from "@langchain/openai"
import { KeyScheduler } from "ai-key-manager"

export async function askWithLangChain(scheduler, prompt) {
  return scheduler.withRetry({
    execute: async ({ apiKey, provider, model }) => {
      const llm = new ChatOpenAI({
        model,
        apiKey,
        configuration: {
          baseURL: "https://openrouter.ai/api/v1"
        }
      })
      return llm.invoke(prompt)
    }
  })
}

import { generateText } from "ai"
import { createOpenAICompatible } from "@ai-sdk/openai-compatible"
import { KeyScheduler } from "ai-key-manager"

export async function askWithVercelAI(scheduler, prompt) {
  return scheduler.withRetry({
    execute: async ({ apiKey, provider, model }) => {
      const gateway = createOpenAICompatible({
        name: provider,
        apiKey,
        baseURL: "https://ai-gateway.vercel.sh/v1"
      })
      const { text } = await generateText({
        model: gateway(model),
        prompt
      })
      return text
    }
  })
}