o4-mini, o3, and UI/response fix.

[alextheradu]

What features would you like to see added?

I can't use the new o4-mini or o3. Also, when I get responses from ChatGPT, they aren't formatted properly and don't use coding blocks and whatnot.

More details

I can't use the new o4-mini or o3. Also, when I get responses from ChatGPT, they aren't formatted properly and don't use coding blocks and whatnot. This is annoying because I have to copy and paste my response into ChatGPT asking it to format it for me.

Which components are impacted by your request?

UI, Endpoints

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[berry-13]

o4-mini and o3 work. make sure that you are using the latest commit. otherwise, wait for the next release, which will be cut soon. responses not being formatted properly is an issue with those models, not the UI. previous models, such as o3-mini, o1, and o1-mini, also had these formatting issues

[otbu2]

I have wondered about this for a while as well. I have defaulted to the older gpt-4o model for this very reason. The newer gpt-4.1 models (full, mini, nano) seem to work fine as well in v0.7.7.

@berry-13 Out of interest, what is it that makes the output of certain models different? They use a different response structure? Is this perhaps specific to reasoning models?

[berry-13]

I have wondered about this for a while as well. I have defaulted to the older gpt-4o model for this very reason. The newer gpt-4.1 models (full, mini, nano) seem to work fine as well in v0.7.7.

@berry-13 Out of interest, what is it that makes the output of certain models different? They use a different response structure? Is this perhaps specific to reasoning models?

I’m not an AI engineer, but from what I’ve seen, reasoning models probably work a bit differently under the hood compared to regular GPT models. This could be why their output sometimes comes out a little weird. They seem more focused on solving problems than on making everything look nice. I’ve noticed they use way less Markdown naturally compared to the normal GPT models. Also, funny enough, I’ve found the tool call results way more accurate than the code block outputs, but that’s just how it is for now