Why did Azure report more completion tokens than my max_completion_tokens setting

Question

Why did Azure report more completion tokens than my max_completion_tokens setting

Hamza Bchiri 20

Hello,

I’m using Azure OpenAI with the o4-mini reasoning model through the Chat Completions API (2025-01-01-preview). In my request I set:

{

"messages": [

{ "role": "user", "content": "example" }

],

"max_completion_tokens": 10000,

"model": "o4-mini"

}

However, when I check the response usage and the Azure portal, I notice that the completion tokens sometimes exceed this limit.

My concern is that:

I expected completion_tokens to be capped at the max_completion_tokens value I set.

Instead, billing includes thousands of reasoning tokens that are not visible in the output.

This makes it difficult to estimate and control costs, since hidden reasoning tokens are unpredictable.

Questions:

Why are reasoning tokens not constrained by the max_completion_tokens parameter?

Is there a way to cap or limit the number of reasoning tokens consumed by reasoning models like o4-mini?

What is the recommended practice for cost control and billing transparency when using reasoning models in production?

Thank you!My concern is that:

I expected completion_tokens to be capped at the max_completion_tokens value I set.

Instead, billing includes thousands of reasoning tokens that are not visible in the output.

This makes it difficult to estimate and control costs, since hidden reasoning tokens are unpredictable.

Questions:

Why are reasoning tokens not constrained by the max_completion_tokens parameter?

Is there a way to cap or limit the number of reasoning tokens consumed by reasoning models like o4-mini?

What is the recommended practice for cost control and billing transparency when using reasoning models in production?

Thank you!

Anshika Varshney 15,535 Reputation points Microsoft External Staff Moderator

2025-09-03T10:06:32.2233333+00:00

Hello Hamza Bchiri,

Building on the points already shared above, I’d like to add a few clarifications that may help. In Azure, the completion_tokens metric represents both the visible output tokens and the hidden reasoning tokens, and at present, there is no separate breakdown exposed in the API or portal. It’s also important to note that the max_completion_tokens parameter only limits the visible output tokens; currently, there is no way to enforce a strict cap on hidden reasoning tokens.

For cost control, in addition to adjusting the reasoning_effort parameter, you can manage usage by keeping prompts shorter and more focused, choosing smaller models (for example, o1-mini instead of o4-mini) when possible, and running load tests with representative prompts to better understand average reasoning token usage. Since reasoning token consumption can vary significantly depending on prompt complexity, it is also a best practice to set up cost alerts in Azure Cost Management so you can monitor usage and prevent unexpected charges.

If the information is useful, please accept the answer and upvote it to assist other community members.

Thankyou!

Answer accepted by question author

Anshika Varshney 15,535 Microsoft External Staff Moderator

Hello Hamza Bchiri,

Thank you for the response.

You’re correct in your understanding, completion_tokens includes both the visible output tokens and the hidden reasoning tokens, while max_completion_tokens only applies to the visible output. The hidden reasoning process isn’t capped, and the only way to influence its cost is through the reasoning_effort parameter, along with prompt design and model selection.

Right now, Azure doesn’t provide a separate breakdown of reasoning vs. output tokens in the usage or billing exports. This means that while you can see the total token usage, you can’t directly distinguish how many were consumed by hidden reasoning.

Please do not forget to "Accept the answer” and “up-vote” wherever the information provided helps you, this can be beneficial to other community members.

Thankyou!

0 comments

1 additional answer

Your answer

Anshika Varshney 15,535 Reputation points Microsoft External Staff Moderator

2025-09-03T10:06:32.2233333+00:00

Hello Hamza Bchiri,

Building on the points already shared above, I’d like to add a few clarifications that may help. In Azure, the completion_tokens metric represents both the visible output tokens and the hidden reasoning tokens, and at present, there is no separate breakdown exposed in the API or portal. It’s also important to note that the max_completion_tokens parameter only limits the visible output tokens; currently, there is no way to enforce a strict cap on hidden reasoning tokens.

For cost control, in addition to adjusting the reasoning_effort parameter, you can manage usage by keeping prompts shorter and more focused, choosing smaller models (for example, o1-mini instead of o4-mini) when possible, and running load tests with representative prompts to better understand average reasoning token usage. Since reasoning token consumption can vary significantly depending on prompt complexity, it is also a best practice to set up cost alerts in Azure Cost Management so you can monitor usage and prevent unexpected charges.

If the information is useful, please accept the answer and upvote it to assist other community members.

Thankyou!

Answer 1

Gowtham CP 8,120 Volunteer Moderator

Hello Hamza Bchiri,

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

There are a few important points regarding your observation with max_completion_tokens when using reasoning models such as o4-mini:

Reasoning tokens are billed as part of completion usage – Reasoning models generate hidden reasoning_tokens during internal reasoning. These tokens do not appear in the response but are included in billing. Reference: Reasoning models in Azure OpenAI

max_completion_tokens applies only to output tokens – The parameter caps the number of tokens in the visible model output, but reasoning tokens consumed internally are not constrained by this limit. Reference: Chat Completions API parameters

Control reasoning cost with reasoning_effort – You can use the reasoning_effort setting (low, medium, or high) to influence how much internal reasoning the model performs, which impacts the number of hidden reasoning tokens. Reference: Reasoning models – control reasoning effort

Monitor usage in Azure Portal – For billing transparency, rely on Azure portal metrics such as processed_prompt_tokens and generated_completion_tokens, and configure alerts through Azure Cost Management. Reference: Manage costs and usage in Azure OpenAI

I hope this helps clarify the behavior and how you can better manage cost.

If the information is useful, please accept the answer and upvote it to assist other community members.

Anshika Varshney 15,535 Reputation points Microsoft External Staff Moderator

2025-09-03T09:02:13.54+00:00

Hello Hamza Bchiri,

Building on the points already shared above, I’d like to add a few clarifications that may help. In Azure, the completion_tokens metric represents both the visible output tokens and the hidden reasoning tokens, and at present, there is no separate breakdown exposed in the API or portal. It’s also important to note that the max_completion_tokens parameter only limits the visible output tokens; currently, there is no way to enforce a strict cap on hidden reasoning tokens.

For cost control, in addition to adjusting the reasoning_effort parameter, you can manage usage by keeping prompts shorter and more focused, choosing smaller models (for example, o1-mini instead of o4-mini) when possible, and running load tests with representative prompts to better understand average reasoning token usage. Since reasoning token consumption can vary significantly depending on prompt complexity, it is also a best practice to set up cost alerts in Azure Cost Management so you can monitor usage and prevent unexpected charges.

If the information is useful, please accept the answer and upvote it to assist other community members.

Thankyou!
Hamza Bchiri 20 Reputation points

2025-09-03T13:23:42.7833333+00:00

Thank you for the clarification.

Just to make sure I fully understand:

completion_tokens in the usage response includes both visible output tokens and hidden reasoning_tokens, which means the billing can be significantly higher than what we see in the response text.

max_completion_tokens only limits the final output, not the hidden reasoning process.

The only way to reduce hidden reasoning costs is by adjusting reasoning_effort.

My concern is around billing transparency for customers. Since reasoning tokens are hidden and not controllable by max_completion_tokens, we have no direct way to estimate or predict the total tokens (and therefore costs) before running the request. This makes it challenging when we need to bill back usage costs to our clients in a clear and auditable way.

Can Microsoft provide more granular usage reporting (e.g., separating visible completion tokens from reasoning tokens in billing exports/portal) so that customers can transparently allocate costs? This would help service providers like us pass through costs with confidence.