Prompt Caching Is a Core GPT-5.6 Feature. Why Are Customers Still Reverse-Engineering It?

API Bugs
1

This is honestly embarrassing.

We’ve spent the last few days doing the engineering work ourselves because there has been no clear communication from either OpenAI or Microsoft.

We tested GPT-5.6 Sol extensively.

Results:

  • Azure Chat Completions → ~96% prompt cache hit.
  • Azure Responses API → 0 cached tokens.
  • GPT-5.6 explicit prompt caching fields (prompt_cache_options and prompt_cache_breakpoint) are accepted and validated by Azure, but they never produce cache writes or reads.
  • Other developers are reporting similar issues on both Azure and OpenAI.

This isn’t some tiny UI bug. Prompt caching is one of the biggest reasons enterprises adopt GPT-5.6. It directly affects production costs, latency, and whether companies can economically use large codebases and RAG systems.

Even a small startup wouldn’t let a production issue with this level of customer impact sit for days without either fixing it or publishing a clear status update. Engineers would be all over it.

OpenAI and Microsoft are multi-billion-dollar companies with world-class engineering teams. From a customer’s perspective, it’s genuinely hard to understand why a core feature affecting customer bills still has no clear explanation or timeline.

If this is a difficult backend issue, tell us. If you’re investigating it, tell us. If it’s an Azure issue, say it’s an Azure issue. If it’s an OpenAI issue, say it’s an OpenAI issue.

But please don’t leave customers spending real money while they have to reverse-engineer the behavior themselves through experiments.

We’ve already reproduced the issue, collected request IDs, compared Chat Completions vs. Responses API, tested explicit and implicit caching, and verified the behavior. We can provide the evidence if it helps the engineering teams investigate.

We aren’t asking for perfection. We’re asking for transparency, communication, and urgency on a feature that directly impacts customer costs.

2

Hi,

Welcome to the developer community, and thank you for sharing the detailed context. We understand how important prompt caching is for production cost, latency, and large-context workloads.

To help us take a closer look, could you please share the caching configuration used in your OpenAI API request, including any relevant request parameters such as prompt_cache_options, prompt_cache_breakpoint, endpoint type, and whether you are using Chat Completions or the Responses API?

If available, request IDs and a minimal reproducible example would also be very helpful for investigation.

For Azure OpenAI Service-specific behavior, the best channel is the Microsoft Tech Community.

Your details will help the appropriate teams better understand and investigate what you are seeing.