Oh this seems bad, and is fairly easy to reproduce using codex cli. You give it a puzzle prompt that it has to reason about and solve, occasionally it will seemingly short circuit and think for exactly 516 tokens, and return the wrong result. When it ends up using 6000-8000 thinking tokens it returns the correct result.
Maybe some issue with adaptive thinking? Another point for local models I guess, don't have to worry about silent server side changes.
Edit: To follow up, it seems to happen quite often. Out of 10 runs of the exact same prompt, 4/10 had this 516 thinking token issue, and every one of these had the wrong solution. So nearly half the time, 5.5 xhigh could be short circuiting and degrading performance. Granted the sample size is small.
I have a philosophical problem with adaptive thinking. It’s a dumb guess for how much thinking budget to allocate ahead of thinking. At least in the context of LLMs there is probably no way of knowing how much thinking (token generation) is needed. The problem space is infinity vast, similarly of two prompts is not going to help any LLM decide how much thinning is needed. Models already stop thinking before hitting the thinking budget.
Why there is so much effort in making adaptive thinking happen and don’t we train models to produce the end of thinning token better?
Feels like a bandaid. We need models to be trained to do a reasonable amount of reasoning (no pub intended):
reason
estimate remaining uncertainty
continue?
reason more
repeat
I agree, adaptive thinking is a pest and without a minimum thinking budget especially Claude for me currently defaults to not think at all even on max effort.
Sequential-thinking was really a step in the right direction, and works almost exactly how you've described, though when it was popular before the reasoning models and even now when I tried it recently I have never once see it use its branching feature and it tends also to have the RLHF urge to answer something "helpful" quickly instead.
At least there should be a tool call that's the equivalent of saying "wow, this is more complicated than I thought". Humans are also often prone to under-allocating reasoning time and coming to wrong conclusions because their reasoning ends up too shallow. But the best humans are great at mentally mapping the problem space and readjusting on the fly
This is what I do in llm-consortium. An arbiter evaluates the response(s) and decides if more iterations are needed. You can also loop until a minimum confidence threshold, but self-reported confidence isn't a great metric.
Right now we have a LOT of band aids. You want to optimize compute and thinking to a particular problem, sort of like we do. Yes you cannot perfectly predict this but you can do decently well and save a ton of tokens at the cost of this band aid being sort of leaky and gross.
But the larger problem is sound, and the answer is something jointly optimized (idk how they do the routing) but it’s hard to shoehorn it into the current paradigm.
This is preliminary, but it seems like it might somehow be related to the `## Intermediary updates` system prompt that's provided to the model. Seems like it forces the model to stop thinking and return early to provide updates. Removing that entirely makes all runs succeed [1].
I wonder if it's somehow getting confused between what's supposed to be an intermediate update vs the final result.
You still have to worry about misconfigured local models. Even the professionals get it wrong, which is why local model performance is uneven across providers.
But in that case you have nobody but yourself to blame, and you can stabilize things yourself at any time by refraining from making any changes. You won't be surprised by a provider. Honestly? That's not just valuable—it's essential.
I've made a passive workaround: a pair of Codex CLI hooks that detect the truncation from the local session transcript and warn — in the TUI at
turn end, and via a message injected into the model's context on your next message.
You can use this small Python script to display an histogram of `reasoning_output_tokens` in your past Codex sessions. I do see a spike at 516 indeed.
import os, glob, re
import matplotlib.pyplot as plt
vals = []
for f in glob.glob(os.path.expanduser(r"~\.codex") + r"\**\*", recursive=True):
if os.path.isfile(f):
try:
s = open(f, "r", encoding="utf-8", errors="ignore").read()
vals += [int(x) for x in re.findall(r'"reasoning_output_tokens"\s*:\s*(\d+)', s)]
except Exception:
pass
plt.hist(vals, bins=200, range=(0, 5000), weights=[100 / len(vals)] * len(vals))
plt.xlabel("reasoning_output_tokens")
plt.ylabel("%")
plt.show()
I’ve definitely experienced step jumps down in quality on an almost daily basis. I usually used xhigh. The experience of relying on codex’s outstandingly thorough coding earlier in the year has evaporated for me. I’m seeing incredibly stupid implementations intermittently, and have simply switched to Claude until openai takes the issue seriously. As far as i could tell they haven’t taken it seriously for the several months I’ve been personally seeing it.
I've switched 3 months ago to Codex because Claude got incredibly stupid. 6 months ago vice versa. It doesn't matter if you use Codex or Claude. Both will fuck with you at some point. Though Codex probably less.
I have noticed this degradation of 5.5 reliability to what, in my experience, I consider Claude-level of reliability since early June.
My journey dealing with this has been transitioning from 5.5 high to 5.5 xhigh to 5.4 high.
5.4 high has been perfectly reliable for me for the last 3 weeks, and I am happy there.
Occasionally, I run some tasks on 5.5 xhigh to check if it has gone back to being 100% perfectly reliable, but, at this point, I am assuming they are just counting on releasing 5.6 rather than dealing with this reliability issue.
I'm on the same journey but I bought a 3090 and put qwen 3.6 27b on it. It covers some things with better reliability. Obviously it doesn't have the breadth of a large model. If that's even a selling point for large models for coding?
i don't ever believe these issues are technical. They're business decisions to downgrade performance because to fix it means $$$$ and you arn't paying them enough.
When I reviewed the conversations affected by this issue, they did not always align with my feeling of "degraded output".
Some were definitely below par, and I recall having to iterate on the generated code more than I wanted to. However, it is only true for a very small number of conversations.
So we're looking at a small set of affected conversations, and even within that small set, only a few will have degraded output, likely because the model can compensate for the reasoning defect over the long conversation.
I think it might affect real work if part of it requires a lot of thinking, i.e. something similar in nature to a puzzle.
There seems to be something wrong with the "commentary" channel related intermediate updates, maybe the model gets confused about what's an intermediate update vs what's the final answer? [1]
Deja Vu... This looks just like the Claude Code performance regression back in April. I just quit my Claude subscription when that happened and went to Codex.
Now I'm kinda thinking of trying per token for both, using GLM 5.2 on Fireworks for most tasks, shelling out to the big boys only when needed. Not totally confident I'll break even though.
Re per token, I had the same reaction, but given both labs are economically advantaged moving customers to per-token consumption... almost want to avoid this on principle. Even if not intentional, benefitting from a degraded product is not something I want to accept or enable.
More now than ever (since original ChatGPT release), the OSS models and open harnesses (eg Pi) are looking mighty attractive.
If pricing is per-token then in theory the vendor can offer you modes that optimize token usage or quality whereas all-you-can eat encourages vendors to satisfy you just enough to keep paying but the trend is towards lower quality responses.
Right? I also quit Claude Code and switch to Codex over that. Now I’m trying to figure out how I could make an extra $65,000 to never have to be concerned about this nonsense again. I know the economics of using open router etc…
But I’m reminded of ~2008 and the rise of “the cloud” as a marketing term that seemed to me to be a cover for dropping an expectation of rich clients, increasing a companies margins around subscriptions that would chip away at local ownership.
Then I got offput by the zealotry and absolutism around “true FoSS”, told myself I was young and moved on.
And really, a lot of subscription models I kind of can appreciate/ tolerate. Might be irksome but whatever, I get that software is expensive to make and it’s not fair in 2026 to value a yearly upgrade of Photoshop at $200. The capricious UI changes to things that’ve worked for 20 years and they take away say the classic color swatches altogether - silly and dumb.
I can use another professionally necessary tool I pay $200/ mo for, Codex, to whip up a classic swatch plugin.
Is that $200 a fair price for my token usage? I think an extremely heavy month I might’ve used a billion tokens?
But that right there is the problem. They have no idea what, specifically, profitability looks like and are going to be pulling endless levers for … I genuinely have no idea how long - at least through 2030/2032 if we tea leaves their debt obligations?
I don’t want to think about any of that. At all. I don’t want to spend time evaluating model preference and degradation and updating the nuances of how I “speak” to an AI because there’s some mystery backend experiment running on the output I use to produce functional outputs — ie the actual products I get paid to build/ maintain.
AI’s something between a tool and coworking companion, and the capricious “personality” changes due to playing with poorly understood and knobs and levers at the inference level - is maddening. To that end, I want a box in the corner I can point to and know exactly the quality of outputs that no one but myself modifies.
The vibe-assumed claude code performance regression, yep. People should stop expecting consistent performance from non-deterministic systems. There is zero empirical corroboration of performance degredation.
There has been a step change... in the amount of whining and complaining coders exhibit lately.
For me, the encrypted reasoning contents, when looking at the base64 string lengtht, show this effect. However, the server-reported reasoning tokens don't. So I assumed it was part of the encryption and/or obfuscation purely. So I don't think there is a real issue.
This is the biggest downside of GPT; thinking is encrypted, so it's more of a black box than kimi/glm/deepseek. You still get thinking summaries though. It's awkward, but workable.
Already reported (not as thoroughly but still quite detailed) two weeks ago and silently “closed as not planned” (keep in mind that the specific reason might be an artifact of GitHub workflow/UX and not actually the intended reason) without a acknowledgement or a response.
What even is the point of a public-facing bug tracker “for devs, by devs” when this is how reports get treated? Might as well use Apple’s Feedback Reporter that routes to /dev/null instead.
Anyway, I find it near impossible to see how this wasn’t already caught and flagged internally – it’s not a subtle pattern. Certainly they are at the very least collecting and graphing reasoning tokens vs model vs effort” and such an obvious spike at (multiple) single stops (not even distributed over a narrow range) should have been an immediate statistical red flag… which leads me to believe (combined with the fact the previously reported issue was closed without comment) that they’re at least internally aware of this behavior even if it’s not necessarily an intentional side effect of some internal forcing metric.
if these ai companies want to be taken seriously as being productivity tools then they're going to have to stop with these ab tests and forcing unproven features onto everyone. it's bad enough that ais are inherently unpredictable in quality of output, but these kinds of changes just make things worse.
anthropic at least does have a latest and stable channel, as the other day they pushed something irritating that would skip question asking phase if you didn't reply in 60 seconds, and it broke my multi terminal workflow. like I don't know what their product people are thinking when they push this kind of stuff, but it made me switch to stable
But this is model behavior and just a public issue tracker which claude code has just without code? I don’t see how it’s any different than https://github.com/anthropics/claude-code for these issues.
I do appreciate that codex is open source generally, but I don’t think it matters for this class of issue as the model is closed still
Not only that, OpenAI generally doesn't gaslight compared to the misanthropic team especially Boris, who was constantly claiming there is nothing wrong with Claude Code. And OpenAI is generous with resets.
There is nothing called "GPT5.5 Codex" unless I've completely misunderstood OpenAI's product line?
Codex is a harness, while GPT-5.5 is a model. The last codex-branded model was 5.3. Codex as a harness ships as a CLI, a desktop app, and a web product (and I'm not at all sure how similar the underlying harness is between them.)
Is the bug here supposed to be with the CLI harness, or the model? Does it also happen in pi, opencode, etc while running GPT-5.5?
If the issue comments are to be believed it could be related to the codex system prompt itself so likely just the harness. I agree the wording is weird but they appear to be referencing GPT 5.5 in Codex because yeah there isn’t a GPT5.5 Codex model.
That was an article in The Information but it didn't read very well to me, I didn't get the impression the author was enough of a technical expert on how LLMs work to credibly evaluate the claim, which came from an insider rumor: https://www.theinformation.com/newsletters/ai-agenda/openai-...
> OpenAI engineers earlier this month told some colleagues they had figured out a way to more than halve the cost of inference, or running existing models, thanks to some newly-discovered optimizations, according to a person with knowledge of those discussions.
I bet that since this bug has made headlines, there are some panicked engineers at OpenAI desperately trying to figure out how to fix it without undoing their “magic optimisation”.
My understanding of the rumor is that it wasn't OpenAI itself, but one of the post-blip OpenAI breakaway groups (rumoured to be Thinking Machines) who have made a breakthrough and seem to be shopping it to OpenAI. I don't think it has actually been implemented by OpenAI yet.
Interesting. So 516 probably means initial 512 byte buffer and a 4 byte header. Then 516 + 518 = 1034...so another 512 + 4 byte header + 2 bytes for a linked list ref or similar, 1034 + 518 = 1552, etc.
It seems to be an inference engine or agent harness defect/misconfig rather. Not only do the issue details not evidence a willful stealth nerf, they actively suggest otherwise: the root cause is crude, and evidently not particularly stealthy (as it's being reported on by a regular user with independently verifiable, exact details).
I don't find "usual user psychosis" particularly fair or tasteful anyhow. You're not left with much more than subjective judgement and speculation/suspicion when all you have is a magic sink of an API endpoint that ingests your context window then spits back a continuation of it. Even if you have a standardized model test suite, claiming a stealth nerf remains an exercise in mind reading (of the people working there). Model quality can degrade without an explicit intention that way, or a downgrade of the underlying infrastructure, after all.
Being tongue-in-cheek conspiratorial, or even actually entertaining the possibility of a nerf, is no psychosis anyways. Not a fan of this trend of people abusing psychology diagnosis terminology like this. I'm sure there are people who go a step beyond and are overconfident in these judgements, maybe in their case it holds. But then that's a minority, and so what you have then is a hyperboly. Doesn't serve anyone.
"They made the model dumber" on literally the same checkpoint with the same prompt on the same quantization running on the same hardware is a staple of AI complaints.
Users are completely incapable of objectively evaluating model quality over time.
Which makes it all the harder to notice actual "stealth nerfs", misconfigurations or other technical issues. Because "they made the model DUMBER, for REAL this time" is background noise.
Maybe its just bad memory but I feel like 5.3 was the best version in terms of token usage and code quality. 5.5 works better but it just eviscerates tokens.
5.3 was incredibly better than 5.4/5.5. I stuck with it for months after 5.4 was released, and kept testing 5.4/5.5 every now and then but they both were too inconsistent, too rash. I switched to 5.5 a few weeks ago and now regret it, but I am no longer seeing 5.3 as an option to use, only 5.3-Spark, which is trash compared to 5.5.
My first thought would be an adjustment to a reasoning budget parameter (using llama.cpp as my reference) which would lead to these results. But no way to know precisely without an OpenAI statement.
It could be a very dishonest way of scaling to demand during peak hours. I know that some people already scoff in this topic about the subjective nature of perceived performance of models. But the model seemed less smart when US comes online (at least from my testing over the month of May).
On my company blog post from a few weeks ago I felt the need to point this out because it had a perceptively more consistent pattern during those overlap times. Should have saved the session logs for further analysis https://webesque.agency/blog/2026-06-19-llms.html
Isn't the standard to use continuous batching? If they are using continuous batching -- I'm curious why generated token length matters, and why they might be clustering them. If not -- I'm curious why they aren't and what is the tradeoff here.
This "~512 batching" makes me think of things like diffusion or prefill.
If they managed to put together some dirty hack that lets them generate about 512 tokens worth of reasoning in parallel instead of in sequence? That would explain it.
It's funny, they sell you a subscription for frontier models, then over time begin to nerf them rapidly and no one talks about it. Should give me a discount when they reduce reasoning effort silently on the server side!
But on the other hand, I've been using 5.5-high on a daily basis in multithreading workflows, i.e. in parallel. I'm barely exhausting my weekly limits. I can't even Human-as-a-Service fast enough to catch up and read all the plans and implementations it does. So there is that.
I’ve been paying $200/mo for Claude code for, IDK, 9 months?
I am 100% sure that I get far more value from today than I did in December. The models are smarter, the limits are higher. It’s possible there’s some “five steps forward, one step back” going on, but it’s hard to imagine complaining about that step back.
> they sell you a subscription for frontier models, then over time begin to nerf them rapidly and no one talks about it.
People talk about it all the time. Just check some of the dozens of forums where its non-stop complaining about nerfs, limit nerfs, performance issues etc...
Is hard to prove that any downgrade is a effect of being deliberately served a lower class model / lower quant, or whatever. Or the "optimizations" hurting the models performance.
The TOS allows for those service "optimizations", so legally, nobody has a foot to stand upon. Like when OpenAI or was it Anthropic played with the cache, this to free up more server resources, only to later discover that its gutted the long term context behavior, and heavily degraded the models as context grew.
If you want 100% guaranteed the same performance/behavior, you need to run a model yourself (be it rented GPUs online or your own local setup). But its going to cost you a lot more ...
Even without stats i know it went bad. In the pass two month barely can do any good scientific writing lately, which of course rely on reasoning. It just writing for gods sake. And it show how far we are from AGI.
> In the pass two month barely can do any good scientific writing lately, which of course rely on reasoning. It just writing for gods sake
Honestly, I think this is a really cool sentence. Imagine going back to 2021 and telling someone this was a legitimate complaint about a pretty cheap and very prevalent technology in 2026.
This is an intermittent issue, you should still be able to get your work done. 5.5 was released two months ago so perhaps you're using 5.5 wrong and some things that worked in 5.4 require tweaking your prompts?
Well, when 5.5 first came out, it was kind of OK, but now it's almost noticeably worse. It can be done, but requires a lot of effort, just more and more round, and actually, the Gemini Pro on the web (which should be 3.1 Pro) is actually doing a more stable job.
The thing that I ask it to do is like take X and Y paper into Z paragraph --- a not-so-silly model should think of how information in X and Y are related and how they support the whole article to synthesize this sentence in a way that is coherent to the article, but 5.5 now will just copy the stuff without any reasoning about the relation. Of course, this will cost a lot of tokens and will be obvious if not done. One clear indicator is that in a few rounds you can see the length of the article get bloated to 2-3x undesirably long, which is clearly because it is not analyzing/synthesizing the info.
this explains so much why gpt 5.5 has been so bad lately it was really puzzling why it struggled so much where when it first came out it was one shotting stuff totally amazing, i tried the prompt that will tell you if your plan is degraded:
codex exec --json --skip-git-repo-check --ephemeral -s read-only --disable memories -m gpt-5.5 -c model_reasoning_effort=high "Do not use external tools. A black bag contains candies with counts: round apple 7, round peach 9, round watermelon 8; star apple 7, star peach 6, star watermelon 4. Shape is distinguishable by touch before drawing; flavor is not. What is the minimum number of candies to draw to guarantee having apple and peach candies of different shapes, i.e. round apple + star peach or round peach + star apple? Give reasoning and final number. The local project dir is irrelevant for this task, do not consult it. "
1. 516, 24
2. 516, 27
3. 516, 12
4. 516, 21
5. 516, 21
This means that the whole time we've been paying for a product that was silently routing to something completely different and inferior from gpt 5.5
Also I read through the github issues and it seems like they closed a previous issue without addressing it ???!!
whooo boy somebody from OpenAI is getting fired over this if not a class action lawsuit is almost guaranteed at this point.
The correct answer is 29, right? You could draw all the watermelons and all the round pieces before drawing a star piece. So the model never gets it right, but it does when listing the cases exhaustively?
This is preliminary, but it seems like it might somehow be related to the `## Intermediary updates` system prompt that's provided to the model. Seems like it forces the model to stop thinking and return early to provide updates. Removing that entirely makes all runs succeed [1].
I wonder if it's somehow getting confused between what's supposed to be an intermediate update vs the final result.
It's been a month I've been using it as they gave me for free, and I found GPT-5 on Codex quite weird/awful. Even x-high. Then I figured out I should try OMP (Pi), and the experience was much better.
You want a refund because of a problem you weren't even aware of until now? And you don't even really know if your work has been impacted by this problem.
If this really is widespread and degrading performance in 40% of the cases, then if OpenAI simultaneously fixes this bug and releases GPT 5.6 within a day or two, then the sudden boost in capability is going to blow people's hair back.
Personally, I would say very likely, to be honest. I gotta go through this a little more, but I actually use 5.5 codex an obscene amount, and I almost never use it for reasoning anymore. It's not even in the same galaxy as far as actually taking out the thinking and using GPT-5.5 or even Claude and then coming back and giving it the reasoning. Blah blah blah, it's the same model. Well, let me tell you, no, it's not, for several reasons, and the delta on intelligence is pretty staggering.
I'm struggling as well to understand, and I think perhaps they mean they use ChatGPT website with GPT-5.5+reasoning for problem solving, and paste the output into Codex CLI/App. I think they're saying that letting Codex CLI/App problem solve with GPT-5.5 isn't as effective. Essentially that the web harness is superior to the agentic engineering harness for problem solving?
Not sure if I agree, but I do happen to use a fair bit of web harness as well, just because I find it to be much more effective at web search and a different type of reasoning. So I must agree a little or else I wouldn't do that.
I know that these types of comments are not really popular here, but this struck a chord with me because I feel the same. They aren't remotely close.
I have codex right now purely because they gave me a month free of ChatGPT Pro, so I have been using it in between my usage resets with claude. Since it's "free money" for me I have been using it exclusively on xHigh.
One of my most frequent prompts is "hey codex worked on ____, but it didn't quite hit the mark, can we review the work..."
Yes, part of this is normal even within the same model -- you have the highest power model review the work for correctness, refactoring opportunities, and so on, but man I tell you, I don't know what it is about codex, this is obviously one guy's anecdote -- same prompting style, same repository documentation ala MD files, same skills, way different results.
All that to say, maybe the bug report is on to something here, and it can be fixed.
Oh this seems bad, and is fairly easy to reproduce using codex cli. You give it a puzzle prompt that it has to reason about and solve, occasionally it will seemingly short circuit and think for exactly 516 tokens, and return the wrong result. When it ends up using 6000-8000 thinking tokens it returns the correct result.
Maybe some issue with adaptive thinking? Another point for local models I guess, don't have to worry about silent server side changes.
Edit: To follow up, it seems to happen quite often. Out of 10 runs of the exact same prompt, 4/10 had this 516 thinking token issue, and every one of these had the wrong solution. So nearly half the time, 5.5 xhigh could be short circuiting and degrading performance. Granted the sample size is small.
I have a philosophical problem with adaptive thinking. It’s a dumb guess for how much thinking budget to allocate ahead of thinking. At least in the context of LLMs there is probably no way of knowing how much thinking (token generation) is needed. The problem space is infinity vast, similarly of two prompts is not going to help any LLM decide how much thinning is needed. Models already stop thinking before hitting the thinking budget.
Why there is so much effort in making adaptive thinking happen and don’t we train models to produce the end of thinning token better?
Feels like a bandaid. We need models to be trained to do a reasonable amount of reasoning (no pub intended):
I agree, adaptive thinking is a pest and without a minimum thinking budget especially Claude for me currently defaults to not think at all even on max effort.
Sequential-thinking was really a step in the right direction, and works almost exactly how you've described, though when it was popular before the reasoning models and even now when I tried it recently I have never once see it use its branching feature and it tends also to have the RLHF urge to answer something "helpful" quickly instead.
At least there should be a tool call that's the equivalent of saying "wow, this is more complicated than I thought". Humans are also often prone to under-allocating reasoning time and coming to wrong conclusions because their reasoning ends up too shallow. But the best humans are great at mentally mapping the problem space and readjusting on the fly
This is what I do in llm-consortium. An arbiter evaluates the response(s) and decides if more iterations are needed. You can also loop until a minimum confidence threshold, but self-reported confidence isn't a great metric.
Right now we have a LOT of band aids. You want to optimize compute and thinking to a particular problem, sort of like we do. Yes you cannot perfectly predict this but you can do decently well and save a ton of tokens at the cost of this band aid being sort of leaky and gross.
But the larger problem is sound, and the answer is something jointly optimized (idk how they do the routing) but it’s hard to shoehorn it into the current paradigm.
Modern LLMs are nothing but band-aids, starting with the absurd bandwidth of HBM3 RAM that makes them possible.
This is preliminary, but it seems like it might somehow be related to the `## Intermediary updates` system prompt that's provided to the model. Seems like it forces the model to stop thinking and return early to provide updates. Removing that entirely makes all runs succeed [1].
I wonder if it's somehow getting confused between what's supposed to be an intermediate update vs the final result.
[1] https://github.com/openai/codex/issues/30364#issuecomment-48...
You still have to worry about misconfigured local models. Even the professionals get it wrong, which is why local model performance is uneven across providers.
And to add insult to injury, some providers will ride on the good reputation of some local model, selling you a terrible quant instead.
With OpenAI, at least my gpt-5.5 is the same as your gpt-5.5. You can't say that about glm for example.
But in that case you have nobody but yourself to blame, and you can stabilize things yourself at any time by refraining from making any changes. You won't be surprised by a provider. Honestly? That's not just valuable—it's essential.
I've made a passive workaround: a pair of Codex CLI hooks that detect the truncation from the local session transcript and warn — in the TUI at turn end, and via a message injected into the model's context on your next message.
See https://github.com/bentoner/codex-516-hook
I wonder if testing during different time/days show patterns? For example, whether the short circuiting happens more often during workday peak hours.
And people pay for those wasted tokens? If that's the case, it is probably good idea to ask for refunds.
You can use this small Python script to display an histogram of `reasoning_output_tokens` in your past Codex sessions. I do see a spike at 516 indeed.
I’ve definitely experienced step jumps down in quality on an almost daily basis. I usually used xhigh. The experience of relying on codex’s outstandingly thorough coding earlier in the year has evaporated for me. I’m seeing incredibly stupid implementations intermittently, and have simply switched to Claude until openai takes the issue seriously. As far as i could tell they haven’t taken it seriously for the several months I’ve been personally seeing it.
I've switched 3 months ago to Codex because Claude got incredibly stupid. 6 months ago vice versa. It doesn't matter if you use Codex or Claude. Both will fuck with you at some point. Though Codex probably less.
At least OpenAI lets me use my own harness. Having to rely on insane PMs letting Claude Mythos go wild on the codebase has not been going well lately.
If you use GitHub Copilot you can switch between them in the same session if you want.
I have noticed this degradation of 5.5 reliability to what, in my experience, I consider Claude-level of reliability since early June.
My journey dealing with this has been transitioning from 5.5 high to 5.5 xhigh to 5.4 high.
5.4 high has been perfectly reliable for me for the last 3 weeks, and I am happy there.
Occasionally, I run some tasks on 5.5 xhigh to check if it has gone back to being 100% perfectly reliable, but, at this point, I am assuming they are just counting on releasing 5.6 rather than dealing with this reliability issue.
I'm on the same journey but I bought a 3090 and put qwen 3.6 27b on it. It covers some things with better reliability. Obviously it doesn't have the breadth of a large model. If that's even a selling point for large models for coding?
i don't ever believe these issues are technical. They're business decisions to downgrade performance because to fix it means $$$$ and you arn't paying them enough.
Indeed, it looks like my work has suffered from the clustering issue as well:
Created a script for this: https://github.com/thehappybug/codex-reasoning-token-checkWhen I reviewed the conversations affected by this issue, they did not always align with my feeling of "degraded output".
Some were definitely below par, and I recall having to iterate on the generated code more than I wanted to. However, it is only true for a very small number of conversations.
So we're looking at a small set of affected conversations, and even within that small set, only a few will have degraded output, likely because the model can compensate for the reasoning defect over the long conversation.
I think it might affect real work if part of it requires a lot of thinking, i.e. something similar in nature to a puzzle.
There seems to be something wrong with the "commentary" channel related intermediate updates, maybe the model gets confused about what's an intermediate update vs what's the final answer? [1]
[1] https://github.com/openai/codex/issues/30364#issuecomment-48...
Oh cool, another source of LLM nondeterminism. Just what we needed!
Deja Vu... This looks just like the Claude Code performance regression back in April. I just quit my Claude subscription when that happened and went to Codex.
Now I'm kinda thinking of trying per token for both, using GLM 5.2 on Fireworks for most tasks, shelling out to the big boys only when needed. Not totally confident I'll break even though.
Re per token, I had the same reaction, but given both labs are economically advantaged moving customers to per-token consumption... almost want to avoid this on principle. Even if not intentional, benefitting from a degraded product is not something I want to accept or enable.
More now than ever (since original ChatGPT release), the OSS models and open harnesses (eg Pi) are looking mighty attractive.
If pricing is per-token then in theory the vendor can offer you modes that optimize token usage or quality whereas all-you-can eat encourages vendors to satisfy you just enough to keep paying but the trend is towards lower quality responses.
Right? I also quit Claude Code and switch to Codex over that. Now I’m trying to figure out how I could make an extra $65,000 to never have to be concerned about this nonsense again. I know the economics of using open router etc…
But I’m reminded of ~2008 and the rise of “the cloud” as a marketing term that seemed to me to be a cover for dropping an expectation of rich clients, increasing a companies margins around subscriptions that would chip away at local ownership.
Then I got offput by the zealotry and absolutism around “true FoSS”, told myself I was young and moved on.
And really, a lot of subscription models I kind of can appreciate/ tolerate. Might be irksome but whatever, I get that software is expensive to make and it’s not fair in 2026 to value a yearly upgrade of Photoshop at $200. The capricious UI changes to things that’ve worked for 20 years and they take away say the classic color swatches altogether - silly and dumb.
I can use another professionally necessary tool I pay $200/ mo for, Codex, to whip up a classic swatch plugin.
Is that $200 a fair price for my token usage? I think an extremely heavy month I might’ve used a billion tokens?
But that right there is the problem. They have no idea what, specifically, profitability looks like and are going to be pulling endless levers for … I genuinely have no idea how long - at least through 2030/2032 if we tea leaves their debt obligations?
I don’t want to think about any of that. At all. I don’t want to spend time evaluating model preference and degradation and updating the nuances of how I “speak” to an AI because there’s some mystery backend experiment running on the output I use to produce functional outputs — ie the actual products I get paid to build/ maintain.
AI’s something between a tool and coworking companion, and the capricious “personality” changes due to playing with poorly understood and knobs and levers at the inference level - is maddening. To that end, I want a box in the corner I can point to and know exactly the quality of outputs that no one but myself modifies.
Fireworks?
Provides access to AI models for a per-token fee. See OpenRouter, they are one of many.
i believe they are referring to https://fireworks.ai/models/fireworks/glm-5p2
The vibe-assumed claude code performance regression, yep. People should stop expecting consistent performance from non-deterministic systems. There is zero empirical corroboration of performance degredation.
There has been a step change... in the amount of whining and complaining coders exhibit lately.
If you bother to look at the issue instead of whining and complaining, you will see the evidence.
Look, I felt it. I didn't wait for the official apology from Anthropic. I quite before they published that, then felt very vindicated when they did.
For me, the encrypted reasoning contents, when looking at the base64 string lengtht, show this effect. However, the server-reported reasoning tokens don't. So I assumed it was part of the encryption and/or obfuscation purely. So I don't think there is a real issue.
This is the biggest downside of GPT; thinking is encrypted, so it's more of a black box than kimi/glm/deepseek. You still get thinking summaries though. It's awkward, but workable.
Already reported (not as thoroughly but still quite detailed) two weeks ago and silently “closed as not planned” (keep in mind that the specific reason might be an artifact of GitHub workflow/UX and not actually the intended reason) without a acknowledgement or a response.
https://github.com/openai/codex/issues/29353
What even is the point of a public-facing bug tracker “for devs, by devs” when this is how reports get treated? Might as well use Apple’s Feedback Reporter that routes to /dev/null instead.
Anyway, I find it near impossible to see how this wasn’t already caught and flagged internally – it’s not a subtle pattern. Certainly they are at the very least collecting and graphing reasoning tokens vs model vs effort” and such an obvious spike at (multiple) single stops (not even distributed over a narrow range) should have been an immediate statistical red flag… which leads me to believe (combined with the fact the previously reported issue was closed without comment) that they’re at least internally aware of this behavior even if it’s not necessarily an intentional side effect of some internal forcing metric.
if these ai companies want to be taken seriously as being productivity tools then they're going to have to stop with these ab tests and forcing unproven features onto everyone. it's bad enough that ais are inherently unpredictable in quality of output, but these kinds of changes just make things worse.
anthropic at least does have a latest and stable channel, as the other day they pushed something irritating that would skip question asking phase if you didn't reply in 60 seconds, and it broke my multi terminal workflow. like I don't know what their product people are thinking when they push this kind of stuff, but it made me switch to stable
I love that Codex is open source and issues like these can surface/be addressed publicly.
But this is model behavior and just a public issue tracker which claude code has just without code? I don’t see how it’s any different than https://github.com/anthropics/claude-code for these issues.
I do appreciate that codex is open source generally, but I don’t think it matters for this class of issue as the model is closed still
I feel openai in general is much more open and real business like compared to anthropic. They’re just a black box.
Not only that, OpenAI generally doesn't gaslight compared to the misanthropic team especially Boris, who was constantly claiming there is nothing wrong with Claude Code. And OpenAI is generous with resets.
There is nothing called "GPT5.5 Codex" unless I've completely misunderstood OpenAI's product line?
Codex is a harness, while GPT-5.5 is a model. The last codex-branded model was 5.3. Codex as a harness ships as a CLI, a desktop app, and a web product (and I'm not at all sure how similar the underlying harness is between them.)
Is the bug here supposed to be with the CLI harness, or the model? Does it also happen in pi, opencode, etc while running GPT-5.5?
If the issue comments are to be believed it could be related to the codex system prompt itself so likely just the harness. I agree the wording is weird but they appear to be referencing GPT 5.5 in Codex because yeah there isn’t a GPT5.5 Codex model.
I swear some days ago someone here claimed Openai succeeded cutting down their compute cost by half with a breakthrough optimization. So this is it?
That was an article in The Information but it didn't read very well to me, I didn't get the impression the author was enough of a technical expert on how LLMs work to credibly evaluate the claim, which came from an insider rumor: https://www.theinformation.com/newsletters/ai-agenda/openai-...
> OpenAI engineers earlier this month told some colleagues they had figured out a way to more than halve the cost of inference, or running existing models, thanks to some newly-discovered optimizations, according to a person with knowledge of those discussions.
I bet that since this bug has made headlines, there are some panicked engineers at OpenAI desperately trying to figure out how to fix it without undoing their “magic optimisation”.
My understanding of the rumor is that it wasn't OpenAI itself, but one of the post-blip OpenAI breakaway groups (rumoured to be Thinking Machines) who have made a breakthrough and seem to be shopping it to OpenAI. I don't think it has actually been implemented by OpenAI yet.
> reasoning-token clustering at 516/1034/1552
Interesting. So 516 probably means initial 512 byte buffer and a 4 byte header. Then 516 + 518 = 1034...so another 512 + 4 byte header + 2 bytes for a linked list ref or similar, 1034 + 518 = 1552, etc.
A rare case "they made the model dumber" where they actually made the model dumber, instead of the usual user psychosis?
This is the second one in a row now (previous was Anthropic flic in Feb/March)
Nah there have been thousands of “they made the model dumber” false alarms between these two.
It seems to be an inference engine or agent harness defect/misconfig rather. Not only do the issue details not evidence a willful stealth nerf, they actively suggest otherwise: the root cause is crude, and evidently not particularly stealthy (as it's being reported on by a regular user with independently verifiable, exact details).
I don't find "usual user psychosis" particularly fair or tasteful anyhow. You're not left with much more than subjective judgement and speculation/suspicion when all you have is a magic sink of an API endpoint that ingests your context window then spits back a continuation of it. Even if you have a standardized model test suite, claiming a stealth nerf remains an exercise in mind reading (of the people working there). Model quality can degrade without an explicit intention that way, or a downgrade of the underlying infrastructure, after all.
Being tongue-in-cheek conspiratorial, or even actually entertaining the possibility of a nerf, is no psychosis anyways. Not a fan of this trend of people abusing psychology diagnosis terminology like this. I'm sure there are people who go a step beyond and are overconfident in these judgements, maybe in their case it holds. But then that's a minority, and so what you have then is a hyperboly. Doesn't serve anyone.
"They made the model dumber" on literally the same checkpoint with the same prompt on the same quantization running on the same hardware is a staple of AI complaints.
Users are completely incapable of objectively evaluating model quality over time.
Which makes it all the harder to notice actual "stealth nerfs", misconfigurations or other technical issues. Because "they made the model DUMBER, for REAL this time" is background noise.
Maybe its just bad memory but I feel like 5.3 was the best version in terms of token usage and code quality. 5.5 works better but it just eviscerates tokens.
5.3 was incredibly better than 5.4/5.5. I stuck with it for months after 5.4 was released, and kept testing 5.4/5.5 every now and then but they both were too inconsistent, too rash. I switched to 5.5 a few weeks ago and now regret it, but I am no longer seeing 5.3 as an option to use, only 5.3-Spark, which is trash compared to 5.5.
It’s not just you this is also my opinion, 5.3-codex was a fantastic model in terms of balancing output quality and cost.
Cheap and efficient enough I could afford to use it on basically everything unlike 5.5 or Opus, but still pretty good, I preferred it to sonnet
They rendered 5.3 unusable for me a few weeks back. It simply was locking up or answering poorly.
I was wondering WTF was happening.
This was past month:
Clearly they are batching reasoning inference in a few multiples of 512 tokens as a throughput optimization
My first thought would be an adjustment to a reasoning budget parameter (using llama.cpp as my reference) which would lead to these results. But no way to know precisely without an OpenAI statement.
It could be a very dishonest way of scaling to demand during peak hours. I know that some people already scoff in this topic about the subjective nature of perceived performance of models. But the model seemed less smart when US comes online (at least from my testing over the month of May).
On my company blog post from a few weeks ago I felt the need to point this out because it had a perceptively more consistent pattern during those overlap times. Should have saved the session logs for further analysis https://webesque.agency/blog/2026-06-19-llms.html
Isn't the standard to use continuous batching? If they are using continuous batching -- I'm curious why generated token length matters, and why they might be clustering them. If not -- I'm curious why they aren't and what is the tradeoff here.
This "~512 batching" makes me think of things like diffusion or prefill.
If they managed to put together some dirty hack that lets them generate about 512 tokens worth of reasoning in parallel instead of in sequence? That would explain it.
It's funny, they sell you a subscription for frontier models, then over time begin to nerf them rapidly and no one talks about it. Should give me a discount when they reduce reasoning effort silently on the server side!
But on the other hand, I've been using 5.5-high on a daily basis in multithreading workflows, i.e. in parallel. I'm barely exhausting my weekly limits. I can't even Human-as-a-Service fast enough to catch up and read all the plans and implementations it does. So there is that.
I’ve been paying $200/mo for Claude code for, IDK, 9 months?
I am 100% sure that I get far more value from today than I did in December. The models are smarter, the limits are higher. It’s possible there’s some “five steps forward, one step back” going on, but it’s hard to imagine complaining about that step back.
> they sell you a subscription for frontier models, then over time begin to nerf them rapidly and no one talks about it.
People talk about it all the time. Just check some of the dozens of forums where its non-stop complaining about nerfs, limit nerfs, performance issues etc...
Is hard to prove that any downgrade is a effect of being deliberately served a lower class model / lower quant, or whatever. Or the "optimizations" hurting the models performance.
The TOS allows for those service "optimizations", so legally, nobody has a foot to stand upon. Like when OpenAI or was it Anthropic played with the cache, this to free up more server resources, only to later discover that its gutted the long term context behavior, and heavily degraded the models as context grew.
If you want 100% guaranteed the same performance/behavior, you need to run a model yourself (be it rented GPUs online or your own local setup). But its going to cost you a lot more ...
Even without stats i know it went bad. In the pass two month barely can do any good scientific writing lately, which of course rely on reasoning. It just writing for gods sake. And it show how far we are from AGI.
> In the pass two month barely can do any good scientific writing lately, which of course rely on reasoning. It just writing for gods sake
Honestly, I think this is a really cool sentence. Imagine going back to 2021 and telling someone this was a legitimate complaint about a pretty cheap and very prevalent technology in 2026.
This is an intermittent issue, you should still be able to get your work done. 5.5 was released two months ago so perhaps you're using 5.5 wrong and some things that worked in 5.4 require tweaking your prompts?
Well, when 5.5 first came out, it was kind of OK, but now it's almost noticeably worse. It can be done, but requires a lot of effort, just more and more round, and actually, the Gemini Pro on the web (which should be 3.1 Pro) is actually doing a more stable job.
The thing that I ask it to do is like take X and Y paper into Z paragraph --- a not-so-silly model should think of how information in X and Y are related and how they support the whole article to synthesize this sentence in a way that is coherent to the article, but 5.5 now will just copy the stuff without any reasoning about the relation. Of course, this will cost a lot of tokens and will be obvious if not done. One clear indicator is that in a few rounds you can see the length of the article get bloated to 2-3x undesirably long, which is clearly because it is not analyzing/synthesizing the info.
I swear all these ai companies are trying to rob us for more price
I have some bad news: every company is trying to rob you for more price
this explains so much why gpt 5.5 has been so bad lately it was really puzzling why it struggled so much where when it first came out it was one shotting stuff totally amazing, i tried the prompt that will tell you if your plan is degraded:
1. 516, 242. 516, 27
3. 516, 12
4. 516, 21
5. 516, 21
This means that the whole time we've been paying for a product that was silently routing to something completely different and inferior from gpt 5.5
Also I read through the github issues and it seems like they closed a previous issue without addressing it ???!!
whooo boy somebody from OpenAI is getting fired over this if not a class action lawsuit is almost guaranteed at this point.
The correct answer is 29, right? You could draw all the watermelons and all the round pieces before drawing a star piece. So the model never gets it right, but it does when listing the cases exhaustively?
No, the answer is 21. You ignored the "Shape is distinguishable by touch before drawing".
Verified this locally myself. Thanks for the concrete test. I guess it's time to give Claude another try.
This is preliminary, but it seems like it might somehow be related to the `## Intermediary updates` system prompt that's provided to the model. Seems like it forces the model to stop thinking and return early to provide updates. Removing that entirely makes all runs succeed [1].
I wonder if it's somehow getting confused between what's supposed to be an intermediate update vs the final result.
[1] https://github.com/openai/codex/issues/30364#issuecomment-48...
I would switch to Claude if they kept Fable 5 in the sub
I'm also afraid to lose my "spot" if I leave codex and 5.6 is coming out so...
tldr:
GPT-5.5 Codex model exhibits a clustering phenomenon in which reasoning_output_tokens cluster at fixed values spaced 518 apart.
These stuck responses at fixed thresholds are strongly correlated with errors in complex tasks.
Observed phenomenon is specific to GPT-5.5; it is much less prevalent in GPT-5.4 and almost absent in GPT-5.2 and 5.3
Reset!
I'm seeing this issue with 5.4 also.
Sounds like a problem with promoting the drafter.
It's been a month I've been using it as they gave me for free, and I found GPT-5 on Codex quite weird/awful. Even x-high. Then I figured out I should try OMP (Pi), and the experience was much better.
I remember GPT 5.2 Codex being fine...
The good experience I had with GPT-5.5 before made me upgrade to Pro this month. Now I want a refund.
You want a refund because of a problem you weren't even aware of until now? And you don't even really know if your work has been impacted by this problem.
No, the decline in GPT-5.5's performance over the past few weeks is clearly noticeable.
This seems really bad…
Does this affect the Codex app too, or just the Codex CLI tool?
From some of the numbers I'm seeing in the GitHub issue, the codex desktop app has the same 516 spikes. So most likely it is affected.
If this really is widespread and degrading performance in 40% of the cases, then if OpenAI simultaneously fixes this bug and releases GPT 5.6 within a day or two, then the sudden boost in capability is going to blow people's hair back.
Personally, I would say very likely, to be honest. I gotta go through this a little more, but I actually use 5.5 codex an obscene amount, and I almost never use it for reasoning anymore. It's not even in the same galaxy as far as actually taking out the thinking and using GPT-5.5 or even Claude and then coming back and giving it the reasoning. Blah blah blah, it's the same model. Well, let me tell you, no, it's not, for several reasons, and the delta on intelligence is pretty staggering.
Care to explain what you mean by that?
I'm struggling as well to understand, and I think perhaps they mean they use ChatGPT website with GPT-5.5+reasoning for problem solving, and paste the output into Codex CLI/App. I think they're saying that letting Codex CLI/App problem solve with GPT-5.5 isn't as effective. Essentially that the web harness is superior to the agentic engineering harness for problem solving?
Not sure if I agree, but I do happen to use a fair bit of web harness as well, just because I find it to be much more effective at web search and a different type of reasoning. So I must agree a little or else I wouldn't do that.
I know that these types of comments are not really popular here, but this struck a chord with me because I feel the same. They aren't remotely close.
I have codex right now purely because they gave me a month free of ChatGPT Pro, so I have been using it in between my usage resets with claude. Since it's "free money" for me I have been using it exclusively on xHigh.
One of my most frequent prompts is "hey codex worked on ____, but it didn't quite hit the mark, can we review the work..."
Yes, part of this is normal even within the same model -- you have the highest power model review the work for correctness, refactoring opportunities, and so on, but man I tell you, I don't know what it is about codex, this is obviously one guy's anecdote -- same prompting style, same repository documentation ala MD files, same skills, way different results.
All that to say, maybe the bug report is on to something here, and it can be fixed.
What?