Beginning today (May 3, 2025), all models (4o, o3, o4-mini-high, and o4-mini) are, without exception, escaping single $ signs (especially bad for the use case of inline math notation). In the past, some of these models would try to use \[ escaping for inline math, but could be instructed not to do so.
This new “feature” (bug) seems impossible to work around. Even when explicitly asked to mimic e.g., ‘$x$’ or to explicitly add a specific number of backslashes, the thought process still prefaces each $ with a co-joined backslash. Even in non-latex output, all single $ signs are escaped (e.g., output like: “The apple costs \$1.10”).
Note that for the purposes of the post, I have intentionally added backslashes myself so the post is readable.
This applies whether logged in or logged out.
This bug seems brand new today; I’ve never seen it before, but it’s a nightmare.
As a note, copying gpt responses from yesterday to the beginning of my chat history does not show any verifiable instances of this happening before.
This is significantly affecting my user experience as well. It makes long math explainers very difficult to read and I need to resort to copying the response into a markdown editor, remove all escape backslashes and render the file myself. it feels like I am the assistant In addition, in canvas mode there is also no way to remove escaped characters and replace them with non-escaped versions. Consistent math output has been very hit and miss in general. Deep research reports always escaped everything making them unreadable, but at least regular chats used to render mostly fine.
This got fixed for like 2 days where all chats would correctly render inline latex equations, and now it’s back to the broken version that only renders block equations but escapes inline ones. Why is this so inconsistent?
It is still a problem. It might be isolated to single chats. But once the problem occurs in a chat one needs to start a new chat. Even explicitly telling Chat to fix it does not help:
Its still an issue. Its unbelievable. I first noticed it back in may but didnt really thought much of it but now as I am using gpt vor math purposes more its driving me crazy.
I have made specific promopts (even in the system prompt) that prohibts GPT to use any backslach before Dollar Sign and underscore. But still it doesnt work. It always tells me that it has fixed the issue but it happens over and over again.
Back when i had issues with Gemini it at least could track where the error came from and then was able to fix it. But not with GPT.
Its just crazy that this is still an issue even with GPT 5….
In addition, in canvas mode there is also no way to remove escaped characters and replace them with non-escaped versions. Consistent math output has been very hit and miss in general. Deep research reports always escaped everything making them unreadable, but at least regular chats used to render mostly fine.