GPT-5.6 Implicit Prompt Caching and Explicit Prompt Cache API Support in Azure OpenAI

Question

GPT-5.6 Implicit Prompt Caching and Explicit Prompt Cache API Support in Azure OpenAI

WayChou 15

Hello,

I have a couple of questions regarding the GPT-5.6 model series in Azure OpenAI.

1. Implicit Prompt Caching

It appears that Implicit Prompt Caching is not working for GPT-5.6 models on Azure OpenAI.

In addition, the GPT-5.6 model pages on OpenRouter show that the Azure provider has a cache hit rate of 0%, while other providers appear to report non-zero cache hit rates.

QQ20260710-145218

Could you please clarify:

Does Azure OpenAI currently support Implicit Prompt Caching for the GPT-5.6 model series?
If it is supported, are there any prerequisites, deployment settings, API versions, or limitations that need to be considered?

2. Support for the New Prompt Cache API

OpenAI recently introduced new prompt cache API options, including:

"prompt_cache_options": { ... }

and

"prompt_cache_breakpoint": {
  "mode": "explicit"
}

Does Azure OpenAI currently support these new API parameters for GPT-5.6 models?

If not, is there an estimated timeline for when these features will become available in Azure OpenAI?

Thank you for your help.

Thanmayi Godithi 11,555 Reputation points Microsoft External Staff Moderator

2026-07-10T08:48:08.7233333+00:00

Hi WayChou ,

Thank you for reaching out on Microsoft Q&A forum.

We will look into your issue and will get back to you shortly.
SRILAKSHMI C 19,730 Reputation points Microsoft External Staff Moderator

2026-07-10T11:03:34.7233333+00:00
Hello @WayChou

Thank you for reaching out to Microsoft Q&A.

Based on the current Azure OpenAI prompt caching documentation, here's what we can confirm regarding GPT-5.6 on Azure OpenAI.

1. Implicit Prompt Caching for GPT-5.6

Azure OpenAI supports prompt caching (implicit prompt caching) for supported models. This feature is enabled by default and does not require any deployment setting or explicit configuration. A cache hit occurs when requests share an identical prompt prefix (minimum 1,024 tokens), and cache usage can be verified through the cached_tokens field returned in the response usage details.

However, the current Azure documentation lists the models that support prompt caching, and GPT-5.6 is not included in the published support matrix.

Since GPT-5.6 is not currently listed, we cannot confirm that prompt caching is supported for this model on Azure OpenAI. Therefore, the 0% cache hit rate you observed through OpenRouter is consistent with the current published support matrix. That said, OpenRouter is a third-party service, and its reported cache statistics are not official Microsoft telemetry.

If you'd like to validate prompt caching behavior, we recommend testing with one of the documented supported models. Prompt caching requires:

A prompt of at least 1,024 tokens

The first 1,024 tokens to be identical across requests

Cache hits can occur for every additional 128 identical tokens

Cache usage can be verified through the cached_tokens field in the response

Please also note that prompt caches are scoped to an Azure subscription and are automatically cleared after a period of inactivity.

2. Support for the New Prompt Cache API

Regarding the newly introduced OpenAI API parameters:

"prompt_cache_options": { ... }

and

"prompt_cache_breakpoint": { "mode": "explicit" }

the current Azure OpenAI documentation does not list these parameters as supported.

The Azure documentation currently documents Azure-specific prompt caching capabilities, including features such as:

prompt_cache_retention

cached_tokens

Prompt caching support beginning with API version 2024-10-01-preview

However, there is no documentation confirming support for prompt_cache_options or prompt_cache_breakpoint in Azure OpenAI.

3. Timeline

At this time, Microsoft has not announced an estimated timeline for supporting these newer explicit prompt caching APIs in Azure OpenAI.

As Azure OpenAI is a managed Azure service, new OpenAI API capabilities are introduced after Azure validation and integration. Feature availability and supported request parameters are announced through the official Azure documentation and release notes when they become available.

Please refer this

Prompt caching in Azure OpenAI https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching?utm_source=chatgpt.com#extended-prompt-cache-retention

Azure OpenAI API version lifecycle https://learn.microsoft.com/azure/ai-services/openai/api-version-lifecycle

Please do let me know if you have further queries.

Thank you.
Bill 0 Reputation points

2026-07-10T13:25:38.66+00:00

There any update on this? Looks like token caching is completely broken against the 5.6 family of models. This isn't about the new implicit/explicit feature.

there is just no read/write caching happening, where for the exact same response API request if you switch back to 5.5 you get proper token caching.

The openrouter screenshot above shows that azure is basically running 5x more expensive for input tokens as the effective price compared to openAI.

Is this not going to get resolved soon? I don't know why anyone would use foundry for 5.6 without token caching working, or why MS would release it without that critical feature.
Ola 20 Reputation points

2026-07-10T13:28:51.35+00:00
Same issue here. We run an internal gateway routing to Azure OpenAI via the Responses API, and GPT-5.6 (terra, sol, luna deployments, West Europe) returns cached_tokens: 0 on every request, including repeated identical prompts with the same prompt_cache_key.

GPT-5.5 on the same resource, same api-version (2025-04-01-preview), same request shape, caches ~90% of input tokens on follow-up turns with no extra config.

Tried every combination on GPT-5.6:

prompt_cache_key alone

prompt_cache_options: {mode: "implicit"}

prompt_cache_options: {mode: "explicit"}

prompt_cache_breakpoint: {mode: "explicit"} on content blocks

All return input_tokens_details.cached_tokens: 0. The field exists in the response, so the plumbing's there — it's just not engaging.

Also ruled out our gateway: same result hitting Azure directly with raw calls, no proxy in the path.

We've pulled GPT-5.6 from the gateway for now and pointed users back at 5.5. Any word on when caching lands for GPT-5.6 on Azure? At 1.25x cache-write pricing, GPT-5.6 without caching is flatly more expensive than 5.5 with it.

Edit: found the root cause. Azure's Responses API doesn't cache GPT-5.6; Chat Completions does.

Same prompt, same GPT-5.6 deployment, same prompt_cache_key, both endpoints:

/openai/v1/responses → cached_tokens: 0 every request

/openai/v1/chat/completions → cached_tokens: 3840 on the second request

Both the GA v1 API (/openai/v1/responses) and the preview API (/openai/responses?api-version=2025-04-01-preview) show the same zero-caching behavior. GPT-5.5 caches fine on both.

Looks like an Azure-side bug — the caching layer for GPT-5.6 isn't wired up on the Responses API specifically. Switching our GPT-5.6 deployments to Chat Completions fixes it on our end.
Satti Shanmukha Sai Reddy 0 Reputation points

2026-07-10T19:30:06.3466667+00:00

@Ola its working
Satti Shanmukha Sai Reddy 0 Reputation points

2026-07-10T19:32:14.0666667+00:00
@Ola
Hi Ola, I think I can confirm your findings with some additional testing.

I configured Continue to use my Azure GPT-5.6 Sol deployment via Chat Completions instead of Responses and tested with repeated prompts (>11K tokens).

Results:

Chat Completions (/openai/deployments/.../chat/completions?api-version=2025-04-01-preview)

First request: cached_tokens = 0

Second request: cached_tokens = 11008

~95.8% cache hit

Responses API (/openai/v1/responses)

First request: cached_tokens = 0

Second request: cached_tokens = 0

I also tested the new GPT-5.6 explicit prompt caching fields (prompt_cache_options and prompt_cache_breakpoint).

Azure accepts and validates those fields (invalid values correctly return HTTP 400), but on valid requests I still observed:

cached_tokens = 0

no cache_write_tokens

no cache reads on subsequent identical requests.

So, for my Azure deployment, it looks like:

✅ Chat Completions + implicit caching works.

❌ Responses API implicit caching does not.

❌ Responses API explicit caching is accepted by the API but doesn’t appear to perform cache writes/reads yet.

This seems to point to an Azure Responses API implementation issue rather than a GPT-5.6 model issue. Have you had a chance to test the new explicit caching parameters as well?
Hendrik 40 Reputation points

2026-07-10T20:30:59.5+00:00

Still not working! The problem is that the issue only appears at the Response API. Please fix this, it is unusable for Agentic Coding!
Hendrik 40 Reputation points

2026-07-10T20:33:25.5033333+00:00

We are running an company for LLM Software like Agentic Coding, ChatGPT etc. We have studied the whole implementation for new GPT 5.6 models. Same Software works perfect with OpenAI endpoints. We also tried to set implicit cache, explicit cache, breakpoints, prompt_cache_key and so one. It is broken for all new GPT 5.6 models which makes them unusable. Its 100% a problem at azure site
Hendrik 40 Reputation points

2026-07-10T20:38:20.6066667+00:00

Response API is needed for most modern LLM cases. Sol does not allow to set reasoning effort for example at Chat Completions or if i remember correctly, images cannot be read as well (not sure). Chat Completions is the old way. Its very frustrating that we cannot use this model at the moment. We also tried US location for those models, same issues. They have been implemented wrong on Azure site on several locations. For us this bug has an incident level, because it makes the model nearly unusable for most cases.
YutongTie-9091 54,026 Reputation points Moderator

2026-07-10T22:06:22.89+00:00

Hello @WayChou I want to share that there is an ongoing bug for Prompt Caching feature and the team is working on fixing this issue. I would suggest you try late next week to see if the in-memory prompt caching is not working for you or not. For explicit mode, there is no announcement around it that Azure supported it right now.

I hope this helps.

Regards,

Yutong
Chetan Suri 0 Reputation points

2026-07-12T15:08:32.5966667+00:00

hopefully Azure team would refund gpt 5.6 usage billing due to caching issue at their end.
YutongTie-9091 54,026 Reputation points Moderator

2026-07-13T19:54:36.54+00:00

Just a quick update, this issue should resolve now. Please have a try.

@WayChou @Chetan Suri @Hendrik @Ola @Bill

Regards,

Yutong
Adithyan Ilangovan 0 Reputation points

2026-07-15T14:14:20.2+00:00

@YutongTie-9091 Same issue and still does not work.
Adithyan Ilangovan 0 Reputation points

2026-07-15T15:25:34.7166667+00:00
Okay did some more tests. Here is the results:

Luna + Responses: 0 cached tokens on identical repeats

Luna + Chat Completions: 0 cached tokens on identical repeats

Sol and Terra: caching worked on both endpoints So looks like Luna is having issue.
Semidark 0 Reputation points

2026-07-15T20:01:21.42+00:00

Hi, I stumbled over this caching issue and was wondering if it is resolved now?
Rusty Miller 5 Reputation points

2026-07-22T02:05:53.87+00:00

Does anyone from Microsoft have an update on this issue being fixed?

Best Regards,
Rusty
Nithish Addepalli 5 Reputation points

2026-07-23T18:42:21.93+00:00

@SRILAKSHMI C Just checking back, is there any update on this?

1 answer

Your answer

Thanmayi Godithi 11,555 Reputation points Microsoft External Staff Moderator

2026-07-10T08:48:08.7233333+00:00

Hi WayChou ,

Thank you for reaching out on Microsoft Q&A forum.

We will look into your issue and will get back to you shortly.
Bill 0 Reputation points

2026-07-10T13:25:38.66+00:00

There any update on this? Looks like token caching is completely broken against the 5.6 family of models. This isn't about the new implicit/explicit feature.

there is just no read/write caching happening, where for the exact same response API request if you switch back to 5.5 you get proper token caching.

The openrouter screenshot above shows that azure is basically running 5x more expensive for input tokens as the effective price compared to openAI.

Is this not going to get resolved soon? I don't know why anyone would use foundry for 5.6 without token caching working, or why MS would release it without that critical feature.
Ola 20 Reputation points

2026-07-10T13:28:51.35+00:00

Same issue here. We run an internal gateway routing to Azure OpenAI via the Responses API, and GPT-5.6 (terra, sol, luna deployments, West Europe) returns cached_tokens: 0 on every request, including repeated identical prompts with the same prompt_cache_key.

GPT-5.5 on the same resource, same api-version (2025-04-01-preview), same request shape, caches ~90% of input tokens on follow-up turns with no extra config.

Tried every combination on GPT-5.6:

prompt_cache_key alone

prompt_cache_options: {mode: "implicit"}

prompt_cache_options: {mode: "explicit"}

prompt_cache_breakpoint: {mode: "explicit"} on content blocks

All return input_tokens_details.cached_tokens: 0. The field exists in the response, so the plumbing's there — it's just not engaging.

Also ruled out our gateway: same result hitting Azure directly with raw calls, no proxy in the path.

We've pulled GPT-5.6 from the gateway for now and pointed users back at 5.5. Any word on when caching lands for GPT-5.6 on Azure? At 1.25x cache-write pricing, GPT-5.6 without caching is flatly more expensive than 5.5 with it.

Edit: found the root cause. Azure's Responses API doesn't cache GPT-5.6; Chat Completions does.

Same prompt, same GPT-5.6 deployment, same prompt_cache_key, both endpoints:

/openai/v1/responses → cached_tokens: 0 every request

/openai/v1/chat/completions → cached_tokens: 3840 on the second request

Both the GA v1 API (/openai/v1/responses) and the preview API (/openai/responses?api-version=2025-04-01-preview) show the same zero-caching behavior. GPT-5.5 caches fine on both.

Looks like an Azure-side bug — the caching layer for GPT-5.6 isn't wired up on the Responses API specifically. Switching our GPT-5.6 deployments to Chat Completions fixes it on our end.
Satti Shanmukha Sai Reddy 0 Reputation points

2026-07-10T19:30:06.3466667+00:00

@Ola its working
Satti Shanmukha Sai Reddy 0 Reputation points

2026-07-10T19:32:14.0666667+00:00

@Ola
Hi Ola, I think I can confirm your findings with some additional testing.

I configured Continue to use my Azure GPT-5.6 Sol deployment via Chat Completions instead of Responses and tested with repeated prompts (>11K tokens).

Results:

Chat Completions (/openai/deployments/.../chat/completions?api-version=2025-04-01-preview)

First request: cached_tokens = 0

Second request: cached_tokens = 11008

~95.8% cache hit

Responses API (/openai/v1/responses)

First request: cached_tokens = 0

Second request: cached_tokens = 0

I also tested the new GPT-5.6 explicit prompt caching fields (prompt_cache_options and prompt_cache_breakpoint).

Azure accepts and validates those fields (invalid values correctly return HTTP 400), but on valid requests I still observed:

cached_tokens = 0

no cache_write_tokens

no cache reads on subsequent identical requests.

So, for my Azure deployment, it looks like:

✅ Chat Completions + implicit caching works.

❌ Responses API implicit caching does not.

❌ Responses API explicit caching is accepted by the API but doesn’t appear to perform cache writes/reads yet.

This seems to point to an Azure Responses API implementation issue rather than a GPT-5.6 model issue. Have you had a chance to test the new explicit caching parameters as well?
Hendrik 40 Reputation points

2026-07-10T20:30:59.5+00:00

Still not working! The problem is that the issue only appears at the Response API. Please fix this, it is unusable for Agentic Coding!
Hendrik 40 Reputation points

2026-07-10T20:33:25.5033333+00:00

We are running an company for LLM Software like Agentic Coding, ChatGPT etc. We have studied the whole implementation for new GPT 5.6 models. Same Software works perfect with OpenAI endpoints. We also tried to set implicit cache, explicit cache, breakpoints, prompt_cache_key and so one. It is broken for all new GPT 5.6 models which makes them unusable. Its 100% a problem at azure site
Hendrik 40 Reputation points

2026-07-10T20:38:20.6066667+00:00

Response API is needed for most modern LLM cases. Sol does not allow to set reasoning effort for example at Chat Completions or if i remember correctly, images cannot be read as well (not sure). Chat Completions is the old way. Its very frustrating that we cannot use this model at the moment. We also tried US location for those models, same issues. They have been implemented wrong on Azure site on several locations. For us this bug has an incident level, because it makes the model nearly unusable for most cases.
YutongTie-9091 54,026 Reputation points Moderator

2026-07-10T22:06:22.89+00:00

Hello @WayChou I want to share that there is an ongoing bug for Prompt Caching feature and the team is working on fixing this issue. I would suggest you try late next week to see if the in-memory prompt caching is not working for you or not. For explicit mode, there is no announcement around it that Azure supported it right now.

I hope this helps.

Regards,

Yutong
Chetan Suri 0 Reputation points

2026-07-12T15:08:32.5966667+00:00

hopefully Azure team would refund gpt 5.6 usage billing due to caching issue at their end.
YutongTie-9091 54,026 Reputation points Moderator

2026-07-13T19:54:36.54+00:00

Just a quick update, this issue should resolve now. Please have a try.

@WayChou @Chetan Suri @Hendrik @Ola @Bill

Regards,

Yutong
Adithyan Ilangovan 0 Reputation points

2026-07-15T14:14:20.2+00:00

@YutongTie-9091 Same issue and still does not work.
Adithyan Ilangovan 0 Reputation points

2026-07-15T15:25:34.7166667+00:00

Okay did some more tests. Here is the results:

Luna + Responses: 0 cached tokens on identical repeats

Luna + Chat Completions: 0 cached tokens on identical repeats

Sol and Terra: caching worked on both endpoints So looks like Luna is having issue.
Semidark 0 Reputation points

2026-07-15T20:01:21.42+00:00

Hi, I stumbled over this caching issue and was wondering if it is resolved now?
Rusty Miller 5 Reputation points

2026-07-22T02:05:53.87+00:00

Does anyone from Microsoft have an update on this issue being fixed?

Best Regards,
Rusty
Nithish Addepalli 5 Reputation points

2026-07-23T18:42:21.93+00:00

@SRILAKSHMI C Just checking back, is there any update on this?

Answer 1

hi WayChou & thx for sharing urs issue here at Q&A portal,

based on what's publicly documented today, I don't think Azure OpenAI has announced support for either of those features for GPT-5.6 yet. Regarding implicit prompt caching, a 0% cache hit rate on OpenRouter isn't enough by itself to conclude that Azure doesn't support it. That metric reflects what the provider exposes through that integration, not necessarily the underlying platform behavior. For the new explicit prompt cache API (prompt_cache_options / prompt_cache_breakpoint), I haven't seen those parameters documented as supported by Azure OpenAI. If they're sent to an Azure endpoint, I'd only rely on features that are explicitly listed in the Azure API documentation. https://learn.microsoft.com/azure/ai-services/openai/reference So at this point, I'd assume those new cache-specific parameters are not supported unless Microsoft confirms otherwise.

As for an ETA, Microsoft generally doesn't publish timelines for when OpenAI API features will appear in Azure OpenAI, so only the product team could comment on that. If anyone from the Azure OpenAI team can confirm whether GPT-5.6 currently supports implicit prompt caching or the new explicit cache API, that would help clear up the confusion, especially since third-party dashboards may not accurately reflect Azure's implementation.

rgds,

Alex

&

If my answer was helpful pls mark it and additional thx if u follow me at Q&A portal

and at my blog https://ctrlaltdel.blog/

Ofer Razon 0 Reputation points

2026-07-23T11:45:39.7733333+00:00

Any news on that???