GPT-5.5 Fast suddenly feels as slow as Standard, with long thinking/context/search stalls
Summary
Over the past two days, GPT-5.5 Fast in Codex has felt significantly slower than before. It now feels closer to the usual Standard speed rather than Fast.
Some simple tasks that would normally complete quickly are now taking 10–20+ minutes, and sometimes even longer, before the task finishes. This feels like a clear regression compared with the previous normal behavior.
What I am observing
- GPT-5.5 Fast feels noticeably slower than it did before.
- Simple tasks can take 10–20+ minutes or longer to complete.
- The “thinking” phase can get stuck for 40+ seconds, sometimes around a minute or more.
- Automatic context compression also appears to take a very long time.
- Reading and searching steps are also extremely slow.
- At the same time, GPT-5.5 Fast also seems less capable / less reliable than before, so the experience feels both slower and less effective.
Why I do not think this is just my local machine or network
I do not believe this is caused by my computer performance. I am using a high-end MacBook Pro configuration:
- MacBook Pro 16-inch
- Apple M5 Max
- 128 GB memory
- macOS Tahoe 26.5
- Codex version 26.519.41501 (3044)
I also do not believe this is simply a network speed issue, because the slowdown is specifically noticeable in Codex/GPT-5.5 Fast behavior and is very different from how it normally performed before.
Expected behavior
GPT-5.5 Fast should feel meaningfully faster than Standard, especially on simple tasks. Thinking, context compression, reading, and searching should not regularly stall for long periods on basic requests.
Actual behavior
GPT-5.5 Fast currently feels like Standard speed or worse, with long stalls during thinking, automatic context compression, reading, and searching.
Request
Could the team please check whether there has been a recent performance regression, routing change, capacity issue, or other backend/client issue affecting GPT-5.5 Fast in Codex?
I am also curious whether other users are seeing the same thing, because this slowdown feels very obvious compared with the previous normal GPT-5.5 Fast experience.
35 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Additional context: I am posting the recent Feedback IDs from several affected Codex conversations. I hope these can help the team locate, compare, and optimize the recent slowdown/regression more quickly.
Feedback IDs:
019e5df0-cb41-7e93-b0bb-b18658e9241d019e5e2b-2f3d-79f1-aa7a-01c8a2825f13019e5aa5-d5c9-7e02-8326-7ebdfea880de019e5bb8-e7fe-7ea1-975c-a461a6eb18c1019e5bbc-bf97-73e3-817b-042febe97482019e5bdf-2b3a-78e1-995c-008444cb5558019e5d7d-5ef0-7942-a913-18ac86391836019e5df8-6b4f-79d2-b05d-a79e11c0fc6d019e5e43-53fb-7543-b760-7611aec7de9bSome of these were from recent simple tasks where GPT-5.5 Fast felt unusually slow, including long stalls during thinking, automatic context compression, reading, or searching.
Yes, I have the same problems.
same problem
same problem
same problem
same problem
same problem
Yup, same
same problem
same problem
This is widespread now, see my issue https://github.com/openai/codex/issues/24260
same problem
Mine is thinking for an hour and could change only 200 lines of code after 1hr 58 minutes
same problem
I encountered the same issue, and it's not just in Fast-mode. The same phenomenon occurs in the Default-mode I use, especially the frequent "thinking" stalls
same problem: it takes me two hours for 1325 lines of code (gpt5.5, xhigh).
same problem
mine did 217 lines after 2 hours and i force stopped it ; output was horrible as it could not even meet a simple objective that could have costed 100 lines of code....
same problem 019e5822-3c39-7612-88b6-8f74b9733985
Interesting update: this only happens on chatgpt codex, if you do it via API then it works fine! This is not a codex bug, it's something deliberate happening to people on coding plans
Yes, it suddenly became about 8x slower than usual. (v26.519.41501 (3044))
Same, and make $100 10x credits plz.
Same problem, I thought I was going crazy. Reddit is filled with threads like this too.
same problem
Same problem
https://github.com/openai/codex/issues/24431#issuecomment-4552839465
Guys you may take a look at my thoughts on the possible root cause of this issue. If you have the environment available, you may also try the approach I suggested.
Interesting that multiple people are independently reporting the same “Fast feels like Standard” pattern over the past few days 👀
What makes this especially notable is that the slowdown doesn’t sound purely inference-related.
A lot of the symptoms seem to cluster around:
Once modern coding agents start mixing:
tool execution + context compression + retrieval + model routing + long-running sessions,
“model latency” becomes surprisingly hard to separate from “runtime scheduling latency”.
In heterogeneous runtimes, users often perceive all orchestration overhead as “the model became slow,” even when the bottleneck may actually sit in routing, continuation scheduling, compaction policy, or backend queue behavior.
We’ve been observing similar patterns while exploring capability-routing / execution-optimization layers:
the more dynamic the runtime becomes, the harder it gets to maintain predictable latency characteristics across sessions 😊
Really useful report.
Same problem, until sunday May 24th works great and monday the speed dramatically decreased. I'm working with 5.4 in the meantime.
Same problem. This becomes quickly problematic for people working with knowledge work and planning projects based on codex performance.
Same problem - A small a module rework is taking more than an hour to complete. :(
also seeing this
Same problem for me.
This feels fixed to me, how is everyone else feeling?
We had a capacity issue last week that resulted in slow responses for some uses. We have mitigated the issue. I'm going to close and lock this issue. If you arrive here in the future because you're experiencing slow performance, look for a recent active issue or open a new one.