Gpt-4.5-preview each request consume 50x premium

Discussion
1

Yesterday, I utilized the GPT-4.5 preview for the first time to troubleshoot a problem in my React project. To my surprise, it burned through 50 premium fast requests without providing a solution. :skull: Ultimately, I turned to GPT-4.1 and managed to resolve the issue with just a single request. :rofl:

This raises a question: why do we have advanced models that incur such high costs but fail to deliver effective outcomes?

2

Oh why would you use 4.5? 4.1 is much better, has higher context and is the latest model of that line.
Not sure why OpenAi makes such problematic naming, 4.5 is way older than 4.1

Some people adjusted their prompts to 4.5 and can use it. But its going to be retired by OpenAI.

OpenAI will retire the GPT-4.5-preview model from the API on July 14, 2025. The model was deprecated in April 2025, with developers notified about its upcoming removal.

3

Thanks for your update. Why we have it in the Cursor then? Shouldn’t it notifies about the cost and being useless?

4

Also note that any preview models are not stable releases. Some people want to try those to test the model or adjust their prompts.

5

Hey. I think you should check usage tab in dashboard (link from new docs)
For me, those “many request scenarios” seem to when I give prompts with many tool calls (like file scanning). It seems to end an API request and then issue a new one after every tool usage and therefore the request costs seems to add up

Would that match your observations?

6

Only MAX mode before Cursor version 0.50 did count each tool call as separate request. This will change as 0.50 is being tested now in beta and rolled out.

GPT-4.5 wasnt available in MAX mode as it is still not listed as supporting that and as it will be removed once not provided by OpenAI.

7

What I meant was: it does not state “tool calls”. But internally, I think a model can only call one tool in their output and what might happen is that this way a human input is - for real - executed with many many requests which is what is now transparent to us.

What I think is cool is that some requests go as low as “0.2 requests cost”. In the linked dashboard you can also see some token counts when hovering over cost cell which made me believe the things I’ve stated