I haven’t seen any other posts about it but for me it seems that running o4-mini in agent mode runs absurdly slowly. Even attaching the files I want to be used, it will grep/read lines 1-200 of tons of even unrelated files and will often just do 25 greps before timing out, requiring a resume prompt. Each grep will also often take 10-30 seconds making each request just run for up to 5 mins or longer.
Am I doing something wrong? I believe o4-mini was supposed to have better agent and tool support but has anyone else experienced something similar/found a solution in using it? It seems to do the best on benchmarks but I have hardly been able to even use it to test for myself so far due to the terrible experience.
Yeah, it’s the same for me. I’m sure they’re aware of it and trying to make improvements, but I’ve had it run for 30+ minutes just to crash . Trying to @ as much relevant code as possible helps so it prevents a bunch of lookups, but overall still slow.
I’ve experienced that as well…it sits there and thinks for a long time, but then it spits out a single line of code that completely fixes the problem I threw at it, so I’m not complaining. Other models (Claude in particular) are super chatty and you get to see every thought they’re having. o4-mini is mysteriously quiet, but seems more capable to get to the bottom of tricky problems without changing a bunch of unrelated things that didn’t need fixing.
Yesterday I had it get stuck in a loop of reading and grepping the same lines of code over and over again. I let it go out of curiosity until it timed out at 25 tool calls. It solved the problem though before it got stuck in the loop, so again, it’s hard to complain.
Today I tried enabling “include project structure” from cursor options, wrote out a detailed project spec doc including file names and mentioned that in my cursor rules and now it seems to waste less time with useless greps but the requests still take a while. I’m not sure which of those steps helped but its slightly better. Still working on testing if its outputs are better than some other models.
Spent some more time with it today. I’m generally liking the outputs. But it is quite slow. It would be nice to see what it’s thinking along the way. Right now it’s just a long list of tool usages followed by a single edit, then a few more tool usages, then finally a terse explanation of what it did.
I believe that openai doesn’t make their thinking tokens visible to api users so that’s on them rather than cursor but I agree it would be very nice.
It seems that some times of day o4 mini will finish in a few minutes but others can take half an hour so maybe it is also an api limit issue? Hopefully they find a way to speed it up a little in the future.
. Trying to @ as much relevant code as possible helps so it prevents a bunch of lookups, but overall still slow.