`gemini-3-pro-preview` returns `undefined` fields inside `providerMetadata`, causing ai-sdk validation failures

[ramnique]

Summary

When calling gemini-3-pro-preview through the OpenRouter provider, the returned providerMetadata object includes fields with the literal value undefined (e.g., cost: undefined). This violates the JSONValue type expected by the Vercel AI SDK.

Reasoning models require passing prior providerMetadata back into subsequent turns. When doing so, ai-sdk performs Zod validation and throws an AI_InvalidPromptError because undefined is not valid JSON and is not allowed in the schema.

Manually sanitizing the metadata (e.g., using a deep prune or JSON.parse(JSON.stringify(...))) works around the issue.

Reproduction

Code

import { openrouter } from "@openrouter/ai-sdk-provider";
import { generateText } from "ai";

const model = openrouter("google/gemini-3-pro-preview");
const messages = [
    { role: "user", content: "Hi" }
];

const { text, providerMetadata } = await generateText({
    model,
    messages,
});
console.log(text);
console.dir(providerMetadata, { depth: null });

const { text: text2 } = await generateText({
    model,
    messages: [
        ...messages,
        { role: "assistant", content: text, providerOptions: providerMetadata },
        { role: "user", content: "wassup?" },
    ],
});
console.log(text2);

Output

Hello! How can I help you today?
{
  openrouter: {
    provider: 'Google AI Studio',
    reasoning_details: [ ...<truncated>... ],
    usage: {
      promptTokens: 2,
      promptTokensDetails: { cachedTokens: 0 },
      completionTokens: 233,
      completionTokensDetails: { reasoningTokens: 224 },
      totalTokens: 235,
      cost: undefined,   <-- ❌ violates schema
      costDetails: { upstreamInferenceCost: 0 }
    }
  }
}
file:///Users/ramnique/work/ai-sdk-undefined/node_modules/ai/dist/index.mjs:1393
    throw new InvalidPromptError2({
          ^

InvalidPromptError [AI_InvalidPromptError]: Invalid prompt: The messages must be a ModelMessage[]. If you have passed a UIMessage[], you can use convertToModelMessages to convert them.
    at standardizePrompt (file:///Users/ramnique/work/ai-sdk-undefined/node_modules/ai/dist/index.mjs:1393:11)
    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
    at async generateText (file:///Users/ramnique/work/ai-sdk-undefined/node_modules/ai/dist/index.mjs:2180:25)
    at async file:///Users/ramnique/work/ai-sdk-undefined/test.js:16:25 {
  cause: _TypeValidationError [AI_TypeValidationError]: Type validation failed: Value: ..<truncated>

see the full log here: https://gist.github.com/ramnique/29a67913be644e9465010607284a5696

Expected Behavior

• providerMetadata should contain only valid JSONValue types:

providerMetadata?: Record<string, Record<string, JSONValue>> | undefined;

The relevant definitions:

• https://github.com/vercel/ai/blob/ai%405.0.104/packages/ai/src/types/provider-metadata.ts
• https://github.com/vercel/ai/blob/ai%405.0.104/packages/ai/src/types/json-value.ts

undefined is not permitted in the above schema

Other info

$ node --version
v22.16.0

  "dependencies": {
    "@openrouter/ai-sdk-provider": "^1.2.8",
    "ai": "^5.0.104"
  }