GPT-5.6 Series - Be careful with your Cost Management! - Incorrect statements/analyses

Question

GPT-5.6 Series - Be careful with your Cost Management! - Incorrect statements/analyses

Hendrik 50

Dear all, if you’ve been using any GPT 5.6 model, please do yourselves a favour and check your usage! Cost management for this model series wasn’t working until yesterday, and Azure has now adjusted the costs without prior notice. Limits and other controls weren’t in place at all until yesterday, and now people are facing a massive cost trap with inaccurate cost analyses!

The absurd thing is: on Reddit, some people are reporting that their cost analysis is showing significantly higher charges for API usage than they should be. We’re affected too. We were charged several hundred euros a day, but we were able to verify our actual usage via the logs, and it’s significantly lower than that (about 6-8 times)!

On average, we have 7–8 percentage of new input tokens, and the rest are cache tokens. Our software is well cache token optimized. Cached tokens should cost slightly more than input tokens. Cache tokens would therefore have to cost a little bit more than the input tokens, as they are 10 times cheaper but account for over 90 percentage of usage.
We’ve been billed over 3000 euros for input tokens and only 300 euros for cached tokens. The only thing they calculated correctly were the output tokens.

Furthermore Azure has charged for cache write tokens, even though they weren’t returned by the APIs at all. For anyone who charges their customers for these, good luck getting that money back.

It’s outrageous what Microsoft is getting up to here! Please contact your billing support team straight away via the Azure Portal and, if possible, check your actual usage before you end up being charged 5–10 times what you’ve actually used.

Hendrik 50 Reputation points

2026-07-20T21:27:36.3533333+00:00

https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20%27Foundry%20Models%27%20and%20contains(meterName,%275.6%20terra%27)%20and%20contains(meterName,%27Std%20DZ%27)&currencyCode=%27EUR%27

Input 15 Euro, Output 2 Euro. Looks good ......... oh my gosh
Adriaan 95 Reputation points

2026-07-21T13:50:36.4766667+00:00

We also fell into this trap and were charged over $5000 for $150 usage on input tokens last week with GPT-5.6-Terra.
After we've already been disputing a similar billing issue with GPT-5.4 that occurred during May with absolutely no recourse thus far. I have little hope that we will be able to successfully dispute these charges either.

So, moral of the story: stay far away from Azure Foundry / Azure OpenAI. Rather create an OpenAI / Anthropic / etc. direct account, as they have real-time billing built around these features. Not this broken batched metered nonsense that Azure billing does.

Or just stay away from AI in general, as it's starting to smell more like scam-central than anything else.

The incompetence / inaction of the Microsoft billing support team and these kinds of mistakes with pricing is absolutely unacceptable.
Hendrik 50 Reputation points

2026-07-22T12:48:20.7566667+00:00

Check this out:

https://www.reddit.com/r/AZURE/comments/1v2kb4d/check_your_azure_openai_bill_we_found_major_gpt54
Phan Hoàng Nam 5 Reputation points

2026-07-24T08:48:51.2066667+00:00

I check the subscription cost every working morning, and this morning I noticed that the bill had suddenly increased from around $3,000 to $9,500. The cost chart was also updated today.

I'm wondering if there is any way to request a billing adjustment. It appears that the cost data for our AI Foundry resource group was only updated after roughly two weeks, instead of every few days as usual. Because of this delay, I had no visibility into our actual usage and couldn't accurately estimate or control our spending in time.

After today's update, I looked at the refreshed cost chart and noticed that it now shows the cost had already increased from about $1,800 to $6,000 last week, and then from $6,000 to $9,500 this week. However, none of these increases were visible until today because the cost data had not been updated. Based on this pattern, I suspect the total may exceed $10,000 once the remaining usage is processed.

Our team's monthly Azure AI spending is normally around $5,000, so this level of cost is far beyond what we expected. I also checked the token usage, and for GPT-5.6-sol alone, the estimated charges are already around $5,000 for input tokens and about $500 for cached tokens.

The main concern is not the usage itself, but the delayed cost reporting. Without timely cost updates, we had no reliable way to monitor our spending or make operational decisions to reduce usage before the costs escalated.
Adriaan 95 Reputation points

2026-07-24T08:54:44.45+00:00

@Phan Hoàng Nam we feel your pain. Please see our Reddit post for info on the Discord server with active Microsoft personnel responding to these issues:

https://www.reddit.com/r/AZURE/comments/1v2kb4d/check_your_azure_openai_bill_we_found_major_gpt54/
https://discord.gg/microsoftfoundry

Right now their advice is to open a support ticket via Azure Portal to track the issue. Link back to the Reddit post in your support ticket so that they know it's linked to the same issue.
Michael BENESTY 10 Reputation points

2026-07-24T08:54:58.6533333+00:00

Same here, cached input are invoiced as non cached input, the reporting showed it, as non cached input is the exact sum of cached write + cache read. We have over 90% cache hit (agentic loop), so invoice level are totally crazy high. No answer to support ticket since 3 days (dev plan, supposed to get something in the next 8 hours).
We are currently working to move away from Azure Foundry. Nothing like that happened on our OpenAI account in 3 years.
Adriaan 95 Reputation points

2026-07-24T09:01:47.9166667+00:00

@Michael BENESTY the worst part is the lack of support. Bugs happen all the time in any system, but the dismissive behaviour and gaslighting we have experienced with the billing and technical support teams assigned to our support tickets are an absolute deal breaker for us. Not to mention that you can expect absolutely zero support from Microsoft with these types of issues if you don't have some form of paid support plan. We are also working to move to provider direct accounts. Microsoft has really dropped the ball on this one.
Michael BENESTY 10 Reputation points

2026-07-24T09:04:37.0633333+00:00

@Adriaan We have a support plan ! Still no respect for their own engagements ...
Adriaan 95 Reputation points

2026-07-24T09:06:50.7633333+00:00

@Michael BENESTY It's ridiculous. It seems like these outsourced support companies Microsoft has employed only care about closing tickets as quickly as possible and have no interest in actually resolving issues.
Hendrik 50 Reputation points

2026-07-25T03:47:36.75+00:00

So far, we have only seen the charges in the cost analysis and usually receive our invoice at the start of the month. We are no longer using the models of the GPT 5.6 Series. We have asked support on several occasions to temporarily suspend the charges, but this has been ignored on multiple occasions. When the amount is debited in this way, the problem had been reported three weeks earlier. We will then arrange for the amount to be refunded. It should be possible for any support team to respond to a publicly known issue and suspend invoicing.

Microsoft’s behaviour in this matter is ridiculous; one might almost call it grossly negligent. We expect a solution next week, hopefully!

To this day Terra’s input and output are still swapped. It’s really beyond belief. It’s as if nobody is taking any notice of the Foundry or this problem. Anywhere else, this would be treated as an incident and resolved within 24 hours, followed by a further 48–72 hours to sort out the billing.
Michaël Benesty 0 Reputation points

2026-07-28T21:04:17.1466667+00:00

Anyone has found a solution?
Hendrik 50 Reputation points

2026-07-29T23:45:46.2566667+00:00

Until now no response. Support still ignoring request to temporarily suspend invoicing. Its pathetic.
Alexander Leeb 0 Reputation points

2026-07-30T21:19:36.4033333+00:00

I can confirm the same behavior on our Sweden Central Data Zone Standard deployments. Our detailed Cost Management export shows that the GPT-5.6 Inp meter contains the same quantities that are also reported and charged under Cd Inp and Cd Wr. All three appear as positive charges, with no offsetting discount or credit.

This makes the effectively billed amount of Inp tokens way too high, cache reads more expensive than ordinary input, cache writes substantially more expensive than the published cache-write rate.

We verified the expected category-exclusive behavior separately through the direct OpenAI API, where total input is partitioned into uncached, cache-read, and cache-write tokens.
Hendrik 50 Reputation points

2026-08-04T11:20:54.1366667+00:00

UPDATE: Microsoft has decided not to refund the money and is instead offering us a laughable amount of credits. All customers are affected. Get ready to hire a good solicitor and find out the best way to block the payment before it goes through. Let’s show Microsoft that their behaviour is illegal and that this won’t do! We recommend switching directly to OpenRouter, Bedrock or OpenAI.
Hendrik 50 Reputation points

2026-09-09T16:29:17.2666667+00:00

We finally got the full amount back! If you still have problems, contact leestott@microsoft.com - He will help you! Very friendly. After contacting him, we was able to resolve the problem in a few days.
Naoto NISHII 0 Reputation points

2026-09-14T10:56:32.09+00:00

Adding a data point from a different model in the same family. gpt-6-astra produced 48.7M input tokens on my subscription and has never generated a meter of any kind — nine days and counting.

What may be useful to this thread: I caught both mechanisms in the act on my own subscription by snapshotting usageDetails daily. On 09-12, payGPrice for the eight gpt-5.6-sol meters dropped 20-33% and the new rates were applied backwards over rows that had already posted, with quantities unchanged. On 09-13, two Grok 4.6 meters appeared dated five days earlier, on a day that previously had no credit-eligible row at all. Row count has gone 106 -> 108 -> 110 over three days while the latest usage date stayed at 09-10.

So "no meter" is demonstrably not "no charge" — the backfill path is live. Full detail here: https://learn.microsoft.com/en-us/answers/questions/6002755/

4 answers

Your answer

Hendrik 50 Reputation points

2026-07-20T21:27:36.3533333+00:00

https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20%27Foundry%20Models%27%20and%20contains(meterName,%275.6%20terra%27)%20and%20contains(meterName,%27Std%20DZ%27)&currencyCode=%27EUR%27

Input 15 Euro, Output 2 Euro. Looks good ......... oh my gosh
Adriaan 95 Reputation points

2026-07-21T13:50:36.4766667+00:00

We also fell into this trap and were charged over $5000 for $150 usage on input tokens last week with GPT-5.6-Terra.
After we've already been disputing a similar billing issue with GPT-5.4 that occurred during May with absolutely no recourse thus far. I have little hope that we will be able to successfully dispute these charges either.

So, moral of the story: stay far away from Azure Foundry / Azure OpenAI. Rather create an OpenAI / Anthropic / etc. direct account, as they have real-time billing built around these features. Not this broken batched metered nonsense that Azure billing does.

Or just stay away from AI in general, as it's starting to smell more like scam-central than anything else.

The incompetence / inaction of the Microsoft billing support team and these kinds of mistakes with pricing is absolutely unacceptable.
Hendrik 50 Reputation points

2026-07-22T12:48:20.7566667+00:00

Check this out:

https://www.reddit.com/r/AZURE/comments/1v2kb4d/check_your_azure_openai_bill_we_found_major_gpt54
Phan Hoàng Nam 5 Reputation points

2026-07-24T08:48:51.2066667+00:00

I check the subscription cost every working morning, and this morning I noticed that the bill had suddenly increased from around $3,000 to $9,500. The cost chart was also updated today.

I'm wondering if there is any way to request a billing adjustment. It appears that the cost data for our AI Foundry resource group was only updated after roughly two weeks, instead of every few days as usual. Because of this delay, I had no visibility into our actual usage and couldn't accurately estimate or control our spending in time.

After today's update, I looked at the refreshed cost chart and noticed that it now shows the cost had already increased from about $1,800 to $6,000 last week, and then from $6,000 to $9,500 this week. However, none of these increases were visible until today because the cost data had not been updated. Based on this pattern, I suspect the total may exceed $10,000 once the remaining usage is processed.

Our team's monthly Azure AI spending is normally around $5,000, so this level of cost is far beyond what we expected. I also checked the token usage, and for GPT-5.6-sol alone, the estimated charges are already around $5,000 for input tokens and about $500 for cached tokens.

The main concern is not the usage itself, but the delayed cost reporting. Without timely cost updates, we had no reliable way to monitor our spending or make operational decisions to reduce usage before the costs escalated.
Adriaan 95 Reputation points

2026-07-24T08:54:44.45+00:00

@Phan Hoàng Nam we feel your pain. Please see our Reddit post for info on the Discord server with active Microsoft personnel responding to these issues:

https://www.reddit.com/r/AZURE/comments/1v2kb4d/check_your_azure_openai_bill_we_found_major_gpt54/
https://discord.gg/microsoftfoundry

Right now their advice is to open a support ticket via Azure Portal to track the issue. Link back to the Reddit post in your support ticket so that they know it's linked to the same issue.
Michael BENESTY 10 Reputation points

2026-07-24T08:54:58.6533333+00:00

Same here, cached input are invoiced as non cached input, the reporting showed it, as non cached input is the exact sum of cached write + cache read. We have over 90% cache hit (agentic loop), so invoice level are totally crazy high. No answer to support ticket since 3 days (dev plan, supposed to get something in the next 8 hours).
We are currently working to move away from Azure Foundry. Nothing like that happened on our OpenAI account in 3 years.
Adriaan 95 Reputation points

2026-07-24T09:01:47.9166667+00:00

@Michael BENESTY the worst part is the lack of support. Bugs happen all the time in any system, but the dismissive behaviour and gaslighting we have experienced with the billing and technical support teams assigned to our support tickets are an absolute deal breaker for us. Not to mention that you can expect absolutely zero support from Microsoft with these types of issues if you don't have some form of paid support plan. We are also working to move to provider direct accounts. Microsoft has really dropped the ball on this one.
Michael BENESTY 10 Reputation points

2026-07-24T09:04:37.0633333+00:00

@Adriaan We have a support plan ! Still no respect for their own engagements ...
Adriaan 95 Reputation points

2026-07-24T09:06:50.7633333+00:00

@Michael BENESTY It's ridiculous. It seems like these outsourced support companies Microsoft has employed only care about closing tickets as quickly as possible and have no interest in actually resolving issues.
Hendrik 50 Reputation points

2026-07-25T03:47:36.75+00:00

So far, we have only seen the charges in the cost analysis and usually receive our invoice at the start of the month. We are no longer using the models of the GPT 5.6 Series. We have asked support on several occasions to temporarily suspend the charges, but this has been ignored on multiple occasions. When the amount is debited in this way, the problem had been reported three weeks earlier. We will then arrange for the amount to be refunded. It should be possible for any support team to respond to a publicly known issue and suspend invoicing.

Microsoft’s behaviour in this matter is ridiculous; one might almost call it grossly negligent. We expect a solution next week, hopefully!

To this day Terra’s input and output are still swapped. It’s really beyond belief. It’s as if nobody is taking any notice of the Foundry or this problem. Anywhere else, this would be treated as an incident and resolved within 24 hours, followed by a further 48–72 hours to sort out the billing.
Michaël Benesty 0 Reputation points

2026-07-28T21:04:17.1466667+00:00

Anyone has found a solution?
Hendrik 50 Reputation points

2026-07-29T23:45:46.2566667+00:00

Until now no response. Support still ignoring request to temporarily suspend invoicing. Its pathetic.
Alexander Leeb 0 Reputation points

2026-07-30T21:19:36.4033333+00:00

I can confirm the same behavior on our Sweden Central Data Zone Standard deployments. Our detailed Cost Management export shows that the GPT-5.6 Inp meter contains the same quantities that are also reported and charged under Cd Inp and Cd Wr. All three appear as positive charges, with no offsetting discount or credit.

This makes the effectively billed amount of Inp tokens way too high, cache reads more expensive than ordinary input, cache writes substantially more expensive than the published cache-write rate.

We verified the expected category-exclusive behavior separately through the direct OpenAI API, where total input is partitioned into uncached, cache-read, and cache-write tokens.
Hendrik 50 Reputation points

2026-08-04T11:20:54.1366667+00:00

UPDATE: Microsoft has decided not to refund the money and is instead offering us a laughable amount of credits. All customers are affected. Get ready to hire a good solicitor and find out the best way to block the payment before it goes through. Let’s show Microsoft that their behaviour is illegal and that this won’t do! We recommend switching directly to OpenRouter, Bedrock or OpenAI.
Hendrik 50 Reputation points

2026-09-09T16:29:17.2666667+00:00

We finally got the full amount back! If you still have problems, contact leestott@microsoft.com - He will help you! Very friendly. After contacting him, we was able to resolve the problem in a few days.
Naoto NISHII 0 Reputation points

2026-09-14T10:56:32.09+00:00

Adding a data point from a different model in the same family. gpt-6-astra produced 48.7M input tokens on my subscription and has never generated a meter of any kind — nine days and counting.

What may be useful to this thread: I caught both mechanisms in the act on my own subscription by snapshotting usageDetails daily. On 09-12, payGPrice for the eight gpt-5.6-sol meters dropped 20-33% and the new rates were applied backwards over rows that had already posted, with quantities unchanged. On 09-13, two Grok 4.6 meters appeared dated five days earlier, on a day that previously had no credit-eligible row at all. Row count has gone 106 -> 108 -> 110 over three days while the latest usage date stayed at 09-10.

So "no meter" is demonstrably not "no charge" — the backfill path is live. Full detail here: https://learn.microsoft.com/en-us/answers/questions/6002755/

Answer 1

Allan Solomon Mejia 9,575

Hello,

Thanks for raising this. It's always a good practice to validate your Azure Cost Management data against your application's API usage logs and Azure Metrics, especially after adopting a new model or pricing tier.

If you observe a significant discrepancy between Azure Cost Analysis, Azure Metrics, and the usage returned by the API, I'd recommend opening a Microsoft Support case and including:

The affected subscription and Azure AI Foundry resource
Model deployment name(s)
Time range of the discrepancy
Azure Metrics and API usage logs
Cost Analysis screenshots or exports

Microsoft recommends using Cost Management together with service metrics to reconcile usage, while the invoice and billing meter records remain the authoritative billing source. If those sources don't align, the product team will need to investigate the metering records.

Please "Accept the Answer" if this information helped you. This will help us and others in the community as well.

Adriaan 95 Reputation points

2026-07-27T06:27:11.2133333+00:00

This has already been raised with the Azure OpenAI team who have admitted the defect and are actively working to resolve it. In addition to incorrect prices being loaded for GPT-5.6-terra and double charging for cached tokens as full-price input tokens, part of the issue was that charges were not reflecting in cost management, even a week after use. So, please, don't post generic answers to complex problems like this. This dismissive behaviour is part of the systemic problem plagueing Microsoft support and why we can't get proper support in the first place.
Xin Jing 10 Reputation points

2026-08-18T18:28:52.84+00:00

You posting a generic answer here is actually worse than saying nothing. People just get more angry since they've already received plenty of generic responses from the support team.

Answer 2

Hendrik 50

Bildschirmfoto vom 2026-08-04 13-19-04

Microsoft has decided not to refund the money and is instead offering us a laughable amount of credits. All customers are affected. Get ready to hire a good solicitor and find out the best way to block the payment before it goes through. Let’s show Microsoft that their behaviour is illegal and that this won’t do! We recommend switching directly to OpenRouter, Bedrock or OpenAI.

Michael BENESTY 10 Reputation points

2026-08-04T14:40:29.34+00:00

Do they recognize there is an issue ? Do the credit value offset the over billed invoiced? We had not heard about them, waiting the invoice right now …
Hendrik 50 Reputation points

2026-08-04T18:07:48.0733333+00:00

Yes at the official Discord channel the problem was confirmed

Answer 3

Xin Jing 10

Has anyone actually gotten a refund, even if it was just as a credit? We were overcharged by $80,000 and still haven’t gotten any real answers. Support just keeps saying they’re investigating, but it’s been a month now and nothing has changed.

I can't believe Azure's support is this bad. I genuinely wish the bankruptcy of Azure would happen sooner.

Xin Jing 10 Reputation points

2026-08-17T18:47:51.84+00:00

Also, the $80K charge exhausted our startup credit. I tried to request a refund through the Azure portal, but it wouldn't let me since those transactions were covered by the startup credit. Then do we just lose all those credits because of Azure's incompetence and super stupid billing mistakes?
Adriaan 95 Reputation points

2026-08-18T07:30:05.3366667+00:00

Nope, unfortunately nothing here either.

We have received communications from a support engineer that further communications will be sent out around credits that will be issued. But that was over a week ago and we have still not seen any such communications.

Also, reaching this point likely cost us more in time spent gathering, notarising, and submitting the required evidence than the value of any credits themselves. In the end, nobody wins here.
Xin Jing 10 Reputation points

2026-08-18T18:25:11.7566667+00:00

Yes, Azure is committing a crime! Our entire service was disrupted because of this, and we had to switch to other providers.
Hendrik 50 Reputation points

2026-08-20T22:27:59.16+00:00

We got a 55% discount on our invoice. Ticket remains open as we still paid to much, even with the discount. Support says everything fine discount was applied automatically, no need. Who wants to start a class action lawsuit? We're joining!
Xin Jing 10 Reputation points

2026-08-20T23:16:04.14+00:00

Same here, is there a group chat for the class action lawsuit, we should start one.

Answer 4

Hendrik 50

We finally got the full amount back! If you still have problems, contact leestott@microsoft.com - He will help you! Very friendly. After contacting him, we was able to resolve the problem in a few days.

0 comments