[feature request]: Support prompt caching parameters

[anpc849]

Bug Description

This is not exactly a bug, but a feature request/infrastructure support request.

I tested prompt caching behavior across different model providers using kaggle-benchmarks. My main finding is that GPT and Qwen appear to benefit from prompt caching, which suggests prompt caching could significantly reduce costs for long-running agent benchmark tasks.

However, other providers I tested, including Anthropic and Google/Gemini, did not appear to benefit from caching in the same way. This may be because provider-specific prompt caching parameters are not currently supported or forwarded through the kaggle-benchmarks model proxy and/or SDK.

This matters because prompt caching can substantially affect both reported cost and latency. Without clear support or visibility into caching behavior, users may incorrectly conclude that one model is faster or cheaper than another, when the difference may actually come from whether prompt caching is being applied.

Kaggle Username

anhoangvo

Execution Environment

Kaggle Notebook

Are you using Kaggle Benchmarks in your code?

Yes — I import and use kaggle_benchmarks directly

Steps to Reproduce

1. Open the shared Kaggle notebook: https://www.kaggle.com/code/anhoangvo/prompt-cache-tester-kaggle-benchmarks
2. Run notebook.
3. Analyze the results.

Code to Reproduce

Expected Behavior

I expected prompt caching support to be consistently available where the underlying provider supports it, either automatically or through provider-specific parameters.

In particular, for providers that require explicit cache parameters, I expected kaggle-benchmarks or its model proxy to preserve and forward those parameters when possible.

Examples of provider-specific caching behavior I found:

• OpenAI / xAI / DeepSeek / Gemini: prompt caching is generally automatic, with cached token information exposed through usage metadata such as usage.prompt_tokens_details.cached_tokens.
• DeepSeek: automatic, but uses fields such as usage.prompt_cache_hit_tokens and usage.prompt_cache_miss_tokens.
• Anthropic: requires explicit cache_control: {"type": "ephemeral"} on the content block; otherwise, caching does not occur.
• Qwen: appears to support cache_control: {"type": "ephemeral"} using Anthropic-like syntax, but only for selected models.
• ZAI / GLM-5: I could not find public caching documentation, so I treated it as automatic.
• xAI: automatic, but recommends using the x-grok-conv-id header per conversation to maximize cache hits.

Actual Behavior

Prompt caching appears to work for some providers/models, such as GPT and Qwen, but not consistently across other providers.

For providers that require explicit caching parameters, such as Anthropic, caching may effectively be disabled if the kaggle-benchmarks model proxy strips, modifies, or fails to forward provider-specific caching fields, such as cache_control.

This makes cost and latency comparisons potentially misleading across providers.

Error Output / Traceback

Environment Info

LLM Model Used

No response

Shared Asset Links

https://www.kaggle.com/code/anhoangvo/prompt-cache-tester-kaggle-benchmarks

Additional Context

No response

Pre-submission Checklist

• I have searched the existing issues and this bug has not been reported before.
• I am using the latest version of kaggle-benchmarks (or have noted my version above).
• I have shared my relevant Kaggle assets (notebooks, datasets, etc.) with kaggle-ai-resources-support@google.com as a Viewer, or I have explained in the issue why I cannot.

[dolaameng]

Thanks @anpc849 ! We are planning for this on server side. And once it's done, we can discuss supporting this in kaggle-benchmarks.

cc @rosbo

[dolaameng]

@calderjo this is tracked in b/524681764. This will also makes our token calculation more accurate.