Azure OpenAI GPT-5.6 Usage Response Missing cache_write_tokens Field

Question

Azure OpenAI GPT-5.6 Usage Response Missing cache_write_tokens Field

EndUser59 10

Azure’s GPT-5.6 series models are accessed through the https://my-resource.openai.azure.com domain. However, the usage object in the response does not include the cache_write_tokens field, which is inconsistent with the official OpenAI API.

What is Azure’s plan for this going forward? Will Azure align with the official OpenAI API behavior, or will it remain different?
Is Azure currently charging for GPT-5.6 cache writes?

Official OpenAI documentation: https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching

Anshika Varshney 15,115 Reputation points Microsoft External Staff Moderator

2026-07-16T20:39:37.93+00:00

Hello @EndUser59

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thankyou!
Anshika Varshney 15,115 Reputation points Microsoft External Staff Moderator

2026-07-19T20:35:52.75+00:00

Hello @EndUser59

If these steps resolve your issue, please mark this answer as Accepted so others can easily find the solution.

If you still face problems, share additional details and we’ll assist further.

Thankyou!
Samuel Makin 0 Reputation points

2026-07-24T14:12:30.1933333+00:00
It is indeed missing from the usage object in the latest v1 responses api (https://xxxxx.openai.azure.com/openai/v1/responses).

This prevents clients from accurately managing costing/usage.

Example:

FIRST_USAGE { "input_tokens": 12610, "input_tokens_details": { "cached_tokens": 0 }, "output_tokens": 16, "output_tokens_details": { "reasoning_tokens": 16 }, "total_tokens": 12626 } SECOND_USAGE { "input_tokens": 12610, "input_tokens_details": { "cached_tokens": 12032 }, "output_tokens": 16, "output_tokens_details": { "reasoning_tokens": 16 }, "total_tokens": 12626 }

With these two requests you can see cache population but no cache_write_tokens in the first request.

Please can this be investigated as a priority?

3 answers

Your answer

Anshika Varshney 15,115 Reputation points Microsoft External Staff Moderator

2026-07-16T20:39:37.93+00:00

Hello @EndUser59

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thankyou!
Anshika Varshney 15,115 Reputation points Microsoft External Staff Moderator

2026-07-19T20:35:52.75+00:00

Hello @EndUser59

If these steps resolve your issue, please mark this answer as Accepted so others can easily find the solution.

If you still face problems, share additional details and we’ll assist further.

Thankyou!
Samuel Makin 0 Reputation points

2026-07-24T14:12:30.1933333+00:00

It is indeed missing from the usage object in the latest v1 responses api (https://xxxxx.openai.azure.com/openai/v1/responses).

This prevents clients from accurately managing costing/usage.

Example:

FIRST_USAGE { "input_tokens": 12610, "input_tokens_details": { "cached_tokens": 0 }, "output_tokens": 16, "output_tokens_details": { "reasoning_tokens": 16 }, "total_tokens": 12626 } SECOND_USAGE { "input_tokens": 12610, "input_tokens_details": { "cached_tokens": 12032 }, "output_tokens": 16, "output_tokens_details": { "reasoning_tokens": 16 }, "total_tokens": 12626 }

With these two requests you can see cache population but no cache_write_tokens in the first request.

Please can this be investigated as a priority?

Answer 1

As far as I know, Azure OpenAI doesn't always expose every usage field that the public OpenAI API returns, so the absence of cache_write_tokens is expected today rather than a bug. If prompt caching is enabled, Azure may still apply caching internally, but the detailed accounting fields can differ from the official OpenAI API.

I haven't seen any public commitment that Azure will add cache_write_tokens or fully align the response schema. If you need definitive information about billing for GPT-5.6 cache writes, it's probably best answered by the Azure OpenAI product team, since pricing and usage reporting are Azure-specific.

Answer 2

Anshika Varshney 15,115 Microsoft External Staff Moderator

Hello @EndUser59

Thank you for raising this question.

Based on the information shared, the cache_write_tokens field is documented for GPT-5.6 models and is expected to appear in the usage response when prompt caching is being utilized.

If you are not seeing cache_write_tokens in the response, it would be helpful to verify:

The exact GPT-5.6 model version being used.
The API version specified in your request.
Whether prompt caching is actually being triggered for the requests being tested.

To investigate further, could you please share:

The API version used in your request (redacting any sensitive information).
A sample usage response showing the missing field.
The model name and deployment type.

This information will help determine whether the behavior is related to API versioning, response formatting, or a potential product issue.

Thankyou!

Kian Xiao 0 Reputation points

2026-07-17T03:14:20.45+00:00
Model versions:

gpt-5.6-sol-2026-07-09

gpt-5.6-luna-2026-07-09

gpt-5.6-terra-2026-07-09

Lifecycle status:

Generally Available (GA)

API:

Azure OpenAI v1 Responses API

没有看到cache_write_tokens

Prompt caching was successfully triggered. For example: { "input_tokens": 6143, "input_tokens_details": { "cached_tokens": 5888 }, "output_tokens": 10, "total_tokens": 6153 }
Anshika Varshney 15,115 Reputation points Microsoft External Staff Moderator

2026-07-19T22:57:22.18+00:00
Hello @Kian Xiao

Since prompt caching is clearly being applied (cached_tokens is populated), the main issue appears to be the absence of the cache_write_tokens field in the usage response rather than a caching failure itself.

To troubleshoot, please verify the following:

Confirm the exact API version being used with the Azure OpenAI v1 Responses API. The cache_write_tokens field may depend on API version support.

Test the same request across all three GPT-5.6 model variants (gpt-5.6-sol, gpt-5.6-luna, and gpt-5.6-terra) and check whether the field is consistently missing.

Compare the first request (which should create/write the cache) with subsequent requests (which should read from the cache). Verify whether cache_write_tokens is absent in both scenarios.

Capture and compare the complete raw API responses, including the usage object, to determine whether the field is omitted by the service or filtered by an SDK/client library.

If the field remains absent across models and API versions, collect:

Model name and version

API version

Sample request/response (sanitized)

Region

Correlation/Request ID

This information would help determine whether this is a documentation mismatch or a product issue requiring further investigation.

Based on the shared response, caching itself appears to be functioning correctly because cached_tokens are being reported, but the missing cache_write_tokens field requires additional validation.

Thankyou!

Answer 3

Per Microsoft's current documentation, this is answered directly: for GPT-5.6 and later model families, cache_write_tokens is a documented, supported field in the usage response — it reports the number of prompt tokens written to cache, and cache writes are billed at 1.25× the uncached input token rate. So to your two questions:

Is Azure charging for GPT-5.6 cache writes? Yes — at 1.25× the standard input token rate, per the current pricing/caching docs.
Will Azure align with OpenAI's API behavior? The field is meant to exist for GPT-5.6+, so if it's genuinely missing from your response, that's likely not an intentional platform difference — more likely an API version issue. Cache-write reporting was added at a specific api-version; if you're on an older pinned version, you'd miss the field even on a model that supports it. Worth double-checking the api-version query param against the latest one in the docs below.

There's also an existing thread on this exact topic worth checking in case it's already been addressed by a Microsoft engineer: GPT-5.6 Implicit Prompt Caching and Explicit Prompt Cache API Support in Azure OpenAI.

Reference: Prompt caching with Azure OpenAI in Microsoft Foundry Models

EndUser59 10 Reputation points

2026-07-14T03:31:44.3666667+00:00

Thank you very much for your answer. I wasn’t aware of the Azure documentation that explains the cache_write_tokens field. Could you provide me with a link to that document?