fix(provider/groq): surface cached_tokens from prompt_tokens_details

[armandmcqueen]

Background

groq now has models with input cache support. For example: https://console.groq.com/docs/model/moonshotai/kimi-k2-instruct-0905

The AI SDK was not correctly extracting the response from groq into the usage object.

Summary

Pull the prompt cache data from the appropriate response field.

Manual Verification

Here is a raw response object from a call with cached input tokens

"rawResponse": {
    "id": "chatcmpl-a46c2c95-4e88-4fca-8165-3bb46e74bb83",
    "object": "chat.completion",
    "created": 1772843463,
    "model": "openai/gpt-oss-20b",
    "choices": [
      {
        "index": 0,
        "message": {
          "role": "assistant",
          "content": "The CAP theorem states that a distributed system can simultaneously provide at most two of the following three guarantees: Consistency, Availability, and Partition tolerance.",
          "reasoning": "We need one sentence answer. Keep concise."
        },
        "logprobs": null,
        "finish_reason": "stop"
      }
    ],
    "usage": {
      "queue_time": 0.041304244,
      "prompt_tokens": 3829,
      "prompt_time": 0.025152422,
      "completion_tokens": 48,
      "completion_time": 0.049817351,
      "total_tokens": 3877,
      "total_time": 0.074969773,
      "prompt_tokens_details": {
        "cached_tokens": 3584
      },
      "completion_tokens_details": {
        "reasoning_tokens": 10
      }
    },
    "usage_breakdown": null,
    "x_groq": {
      "id": "req_xxxx",
      "seed": 1536934315
    },
    "service_tier": "on_demand"
  },

Checklist

• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Future Work

Related Issues