As the title says - on a mathematical-related query without agentic calls at 128K max tokens context for GPT-5.6 Sol on the API Platform (not ChatGPT/Codex) for a single query:
```json
{
“model”: “gpt-5.6-sol”,
“max_output_tokens”: 128000,
“service_tier”: “flex”,
“background”: true,
“reasoning”: {“effort”: “max”, “summary”: “auto”},
“tools”: []
}
```
I was charged for 16M output tokens (reasoning tokens correctly stopped at 128K). What is even more concerning is that the model shouldn’t support that amount of output tokens for a single query at all. This, of course, cost $ 250 and was replicated through several hundred calls on different mathematical questions, racking up 1000s of dollars in costs. I contacted the support, and they are completely unhelpful - I only get LLM answers about checking the documentation of Codex despite explaining 3 times so far that this happened on the API Platform and providing exact information like key, org, response ids etc. I am super frustrated with the support, and I do not know what to do. Additionally, this has been dragging for days now, and we are left holding the bag. Please advise what to do here!
They closed all of my support tickets … I even disputed one and they closed the dispute too. I also send them your ticket last week and nothing… We are a bit out of ideas what to do
I will write here again because the mods are trying to silence me and this is unacceptable. We did see the same issue, as reported by OP himself. We got terrible and frankly irresponsible behaviour by the support which is now continued by the mods here who hid my original message explaining this basic fact. To mods: Yes additional reports of the same incident are relevant information to this tread. And yes this needs to be solved by openai and openai needs to return the money to everyone involved in this accident and no it is not our job to have to chase support for a month to solve this.
No, this was an exactly reproducible issue with a clear and high quality write-up from OP.
Especially the 9x usage difference was exactly reproducible.
It is very different from this topic and that is why your post was off-topic there.
Our behaviour at least at the time was extremely reproducible and caused as we learned later by the same issue. The difference is that our case was an extreme outlier so it produced 250 $ for a single call. All of this was explained to the support though who ignored us repeatedly as explained in the other tread
From what I understand here we are talking about a single call to 5.6-Sol via the API that consumed 16 million tokens (reasoning and output) and was billed accordingly. And this did happen several times.