The thread has little explanation as to what weird thing they’re doing to Codex that is making the default work poorly, and it kind of seems like it’s getting confused about whether it wants to set the caching mode or the breakpoint or both.
In any case, I find the behavior change interesting. It sounds to be like 5.5 and below may have been using a conventional attention scheme where a cached KV sequence can be easily used to restore a prefix of itself, but perhaps 5.6 is using linear attention or LSTM or another recurrent scheme where you cannot rewind the model state by just truncating it.
I have noticed the same thing starting with 5.6 when editing my last prompt inside the vscode codex plugin, I’ve seen the model’s thinking respond to the edit with a remark.
Slightly bummed out about it because in the past you could try different situations during a planning session and it wouldn’t pollute the cache but now it does. I’m not sure if forking the conversation has the same problem.
Then why are they (US frontier models) still so far ahead whenever I test them against the latest Chinese models? No bias here, I'd love them to be better for my own personal gain, but I haven't seen it
Behind on architecture, ahead on training? It seemed pretty obvious to me that the opus 4.7 and 4.8 releases were more about trying to retain 4.6-level capabilities while being cheaper to run, which would fit. And they can burn so much money on training.
I don't know I just care about the end result. And yeah what you're mentioning here is a pretty common conspiracy theory but you don't actually have any insight into that do you?
You didn't answer my question you literally just asked me another question and then parroted a common talking point about Opus models (which isn't even frontier - Fable is)
There is so much misinformation in the ecosystem, parrots just hitting "Reply" without thinking one iota, you really cannot trust "human" opinions on the internet anymore, anywhere.
Same with local LLMs, I'd love to use them for my day-to-day software engineering, and I'm not exactly GPU poor, then people with 12GB VRAM try to convince me their local setup is perfectly fine running latest Qwen and it does real engineering but whenever I try, they're a far cry from what Codex+GPT 5.x would do.
Only way to be sure is creating your own private benchmarks and use those, and the difference in quality becomes very apparent, very quickly, for your specific use cases.
> It sounds to be like 5.5 and below may have been using a conventional attention scheme where a cached KV sequence can be easily used to restore a prefix of itself, but perhaps 5.6 is using linear attention or LSTM or another recurrent scheme where you cannot rewind the model state by just truncating it.
I feel like this is the kind of substantial change to your product that you would need to tell your customers about. It would be simply disrespectful to your customers to not disclose this upfront.
“Users are the product” is a phrase used when the users aren’t the ones paying for a free service. For a paid API the users absolutely are the customers.
> Wow, that whole thread is borderline incoherent, presumably generated by an AI without adequate oversight.
What sucks is that every issue tracker for these agent harnesses are the same, and this shit hides real issues!
For example, Codex started encrypting messages from a agent to the sub-agents when you use Sol + Ultra, which is terrible for debugging for obvious reasons. This GitHub issue exists for this: https://github.com/openai/codex/issues/28058
Fine, the opening issue isn't concise exactly, but it's mostly clear what's going on. After a few messages, someone who uses LLMs without reviewing their output starts participating in the discussion, pastes huge walls of texts completely missing the point and overall just bloating the conversation so now whenever a maintainer actually want to address it, they have to wade through 20+ messages of just pure shit and bloat, to even understand what's going on.
Kind of wish some projects started having forums specifically for people who pay for forum access, or some other gate to get rid of these LLM lowlifers who cannot compose a simple message to explain what's in their head, and instead have to ruin perfectly fine conversations/discussions with their verbal poop.
My most awkward experience was a maintainer commenting on my feature request just to prompt a bot to "explain to issue reporter why this is very hard to implement."
It felt like they were trying to avoid me. They could have simply addressed me and given me the explanation they gave to the bot: it would have been simpler for him and more polite. I did in fact reply without waiting for the bot.
It's as though you're talking to someone and they were said to their 'assistant', "Explain this to this person" and walked away. It doesn't really matter what the explanation is, it's just gross.
The fact that companies with access to SOTA non-public AI keep having these kinds of dumb bugs in something as simple as a chat app helps validate my disbelief of people claiming AI is ready to replace all software development.
> The fact that companies with access to SOTA non-public AI keep having these kinds of dumb bugs in something as simple as a chat app helps validate my disbelief of people claiming AI is ready to replace all software development.
Except leadership doesn't care about dumb bugs. Never has, never will (until it's too late). It cares about velocity and cost.
They'd totally replace all software development with worse AI software development in a heartbeat.
> Except leadership doesn't care about dumb bugs. Never has, never will (until it's too late). It cares about velocity and cost.
This is the very moment at which I started my own business. I prefer to work for myself with some quality standards than to be in a rush in front of a prompt (not that I do not use AI at all, I do, but not for generating code most of the time).
I knew the future, at that time was basically: pressure for speed, taking ownership of course, even if they rush you. Wild-guess, probably with an AI, to add on top more trch debt. Make everything unmaintainable in the long term.
So this was the perfect moment to show that things can be done in another way and quality can be kept higher than the competition bc what I am seeing lately is people throwing things in a rush. Better twopieces of well-crafted software than 10 pieces of unmantainable junk.
Precisely. I think everyone has been affected by the fearmongering and gaslighting to some dfgree. But step back and try and see whether software's getting better as a whole or going into reverse? OpenAI has basically unlimited internal compute and talent yet they screw this up amongst many other things. Shouldn't it be a 5 minute job for someone at AI to spin up a team of agents annd make sure this sort of thing never happens?
At a high level, most uses of AI I've seen seem to be people building other AI tools, orchestrators, managers, agent managers etc. But these are all means to ends. I mean I guess it's nice to play aroud with harnesses and command agents to do this and that, but where are the tangible outputs?
I just see so many people boasting of their token burn and the complexity of their agentic setup, yet they rarely show the actual outputs
Our codex on AWS Bedrock read / write cache ratio was less than 5%. Cache writes are very expensive and they were never being used. This results in codex on Bedrock causing ~10x what it should due to no caching and massive writes.
The workaround in issue resolved for me:
web_search = "disabled"
If you’ve got a workaround, I’d suggest updating the issue description to have it up top there so similarly impacted users can spot it quickly and benefit.
My conspiratorial mind thinks they're doing this deliberately and using the resets to mask things so people can't tell their limits are reduced. The $200 / month plan covers about 2 days of usage for me right now.
Codex usage feels exorbitantly high since today. They [0] are denying it, but the number of anecdotal users who decided to raise this as an issue (as a result it's trending on X) says otherwise.
I love the fact that devs are still complaining that invoices are able to grow from $300 to $1000+
How can anyone use a platform where this is even an issue?
Just because AWS is a failure in this regard, doesn't mean there aren't alternatives with fixed prices or others with easily settable limits.
Prompt edits leaking into the cache and affecting model responses is exactly the kind of billing-relevant behavior change that should be in release notes, not discovered by users.
Rookie mistake - it seems like they didn't follow manufacturers' guidance when installing the 10x engineers. One needs to clearly define which metric should be 10x'd before powering them up.
You know, installing unsloth studio lets you use codex against a local qwen 3.8 instance, which does ~10 tok/sec without GPU on a modern machine, and 100+ tok/sec on a 5090, and is incredibly good.
Applying Occam’s razor, which do you think is more likely:
1. OpenAI intentionally adds random overcharges.
2. OpenAI deprioritizes fixing actual bugs that cause occasional overcharges because doing so won’t affect their bottom line.
Wow, that whole thread is borderline incoherent, presumably generated by an AI without adequate oversight.
Here are the docs:
https://developers.openai.com/api/docs/guides/prompt-caching...
The thread has little explanation as to what weird thing they’re doing to Codex that is making the default work poorly, and it kind of seems like it’s getting confused about whether it wants to set the caching mode or the breakpoint or both.
In any case, I find the behavior change interesting. It sounds to be like 5.5 and below may have been using a conventional attention scheme where a cached KV sequence can be easily used to restore a prefix of itself, but perhaps 5.6 is using linear attention or LSTM or another recurrent scheme where you cannot rewind the model state by just truncating it.
I have noticed the same thing starting with 5.6 when editing my last prompt inside the vscode codex plugin, I’ve seen the model’s thinking respond to the edit with a remark.
Slightly bummed out about it because in the past you could try different situations during a planning session and it wouldn’t pollute the cache but now it does. I’m not sure if forking the conversation has the same problem.
It’s really cool that we have this proof that US companies are half year behind Chinese models in architecture.
Then why are they (US frontier models) still so far ahead whenever I test them against the latest Chinese models? No bias here, I'd love them to be better for my own personal gain, but I haven't seen it
Behind on architecture, ahead on training? It seemed pretty obvious to me that the opus 4.7 and 4.8 releases were more about trying to retain 4.6-level capabilities while being cheaper to run, which would fit. And they can burn so much money on training.
I don't know I just care about the end result. And yeah what you're mentioning here is a pretty common conspiracy theory but you don't actually have any insight into that do you?
Apologies, you asked a question. I assumed that meant you were interested.
You didn't answer my question you literally just asked me another question and then parroted a common talking point about Opus models (which isn't even frontier - Fable is)
Meh. You asked why, I suggested a possible reason, you them said you don't care about why. Congratulations, I now regret engaging.
There is so much misinformation in the ecosystem, parrots just hitting "Reply" without thinking one iota, you really cannot trust "human" opinions on the internet anymore, anywhere.
Same with local LLMs, I'd love to use them for my day-to-day software engineering, and I'm not exactly GPU poor, then people with 12GB VRAM try to convince me their local setup is perfectly fine running latest Qwen and it does real engineering but whenever I try, they're a far cry from what Codex+GPT 5.x would do.
Only way to be sure is creating your own private benchmarks and use those, and the difference in quality becomes very apparent, very quickly, for your specific use cases.
What is the proof?
Nemotron was using hybrid with recurrence via mamba layers since around April 2025.
> It sounds to be like 5.5 and below may have been using a conventional attention scheme where a cached KV sequence can be easily used to restore a prefix of itself, but perhaps 5.6 is using linear attention or LSTM or another recurrent scheme where you cannot rewind the model state by just truncating it.
I feel like this is the kind of substantial change to your product that you would need to tell your customers about. It would be simply disrespectful to your customers to not disclose this upfront.
It depends on who they consider the customers. Shareholders and govt are the customers, not users.
Users is the product.
“Users are the product” is a phrase used when the users aren’t the ones paying for a free service. For a paid API the users absolutely are the customers.
> Wow, that whole thread is borderline incoherent, presumably generated by an AI without adequate oversight.
What sucks is that every issue tracker for these agent harnesses are the same, and this shit hides real issues!
For example, Codex started encrypting messages from a agent to the sub-agents when you use Sol + Ultra, which is terrible for debugging for obvious reasons. This GitHub issue exists for this: https://github.com/openai/codex/issues/28058
Fine, the opening issue isn't concise exactly, but it's mostly clear what's going on. After a few messages, someone who uses LLMs without reviewing their output starts participating in the discussion, pastes huge walls of texts completely missing the point and overall just bloating the conversation so now whenever a maintainer actually want to address it, they have to wade through 20+ messages of just pure shit and bloat, to even understand what's going on.
Kind of wish some projects started having forums specifically for people who pay for forum access, or some other gate to get rid of these LLM lowlifers who cannot compose a simple message to explain what's in their head, and instead have to ruin perfectly fine conversations/discussions with their verbal poop.
i dont know what the hell is going on lately i pop in to issues or discussions and its agents talking to each other or telling me what PR to merge
I mean i use AI too but was taken back when an agent popped up dictating what i should do and so on....felt weird
My most awkward experience was a maintainer commenting on my feature request just to prompt a bot to "explain to issue reporter why this is very hard to implement."
It felt like they were trying to avoid me. They could have simply addressed me and given me the explanation they gave to the bot: it would have been simpler for him and more polite. I did in fact reply without waiting for the bot.
It's as though you're talking to someone and they were said to their 'assistant', "Explain this to this person" and walked away. It doesn't really matter what the explanation is, it's just gross.
The fact that companies with access to SOTA non-public AI keep having these kinds of dumb bugs in something as simple as a chat app helps validate my disbelief of people claiming AI is ready to replace all software development.
> The fact that companies with access to SOTA non-public AI keep having these kinds of dumb bugs in something as simple as a chat app helps validate my disbelief of people claiming AI is ready to replace all software development.
Except leadership doesn't care about dumb bugs. Never has, never will (until it's too late). It cares about velocity and cost.
They'd totally replace all software development with worse AI software development in a heartbeat.
> Except leadership doesn't care about dumb bugs. Never has, never will (until it's too late). It cares about velocity and cost.
This is the very moment at which I started my own business. I prefer to work for myself with some quality standards than to be in a rush in front of a prompt (not that I do not use AI at all, I do, but not for generating code most of the time).
I knew the future, at that time was basically: pressure for speed, taking ownership of course, even if they rush you. Wild-guess, probably with an AI, to add on top more trch debt. Make everything unmaintainable in the long term.
So this was the perfect moment to show that things can be done in another way and quality can be kept higher than the competition bc what I am seeing lately is people throwing things in a rush. Better twopieces of well-crafted software than 10 pieces of unmantainable junk.
I think being self-employed is the only way you can make sure that your work is allowed to reach your own minimal quality standards.
Precisely. I think everyone has been affected by the fearmongering and gaslighting to some dfgree. But step back and try and see whether software's getting better as a whole or going into reverse? OpenAI has basically unlimited internal compute and talent yet they screw this up amongst many other things. Shouldn't it be a 5 minute job for someone at AI to spin up a team of agents annd make sure this sort of thing never happens?
At a high level, most uses of AI I've seen seem to be people building other AI tools, orchestrators, managers, agent managers etc. But these are all means to ends. I mean I guess it's nice to play aroud with harnesses and command agents to do this and that, but where are the tangible outputs?
I just see so many people boasting of their token burn and the complexity of their agentic setup, yet they rarely show the actual outputs
we've had bugs for decades; the people wanting to replace humans arn't going to care much.
AI is ready to replace both jobs and companies.
The companies you see struggling are ripe for disruption.
So OpenAI the company making the AI and using it inhouse is ripe for disruption? What?
I want some of what he’s having please
Our codex on AWS Bedrock read / write cache ratio was less than 5%. Cache writes are very expensive and they were never being used. This results in codex on Bedrock causing ~10x what it should due to no caching and massive writes.
The workaround in issue resolved for me: web_search = "disabled"
If you’ve got a workaround, I’d suggest updating the issue description to have it up top there so similarly impacted users can spot it quickly and benefit.
It's already mentioned in the issue...
It’s in a comment halfway down the page. It’s not in the issue description.
If enough comments are added to the discussion, it might end up being collapsed.
"causing" -> "costing", right?
In this case yeah. If it’s not reading the cache then it has to compute all the context window again and not just the newest tokens.
Way to bury the lede..
Something is wrong with the codex app too, burning usage like crazy lately.
There haven't been any free resets in the past week, there were 4 in the first half of the month
Yeah regardless of comments by the team to the contrary (https://x.com/thsottiaux/status/2090675027670978569) I have observed this in the cdoex app.
My conspiratorial mind thinks they're doing this deliberately and using the resets to mask things so people can't tell their limits are reduced. The $200 / month plan covers about 2 days of usage for me right now.
indeed it has anybody know whats going on at openai ??
Maybe preparing for their IPO?
Codex usage feels exorbitantly high since today. They [0] are denying it, but the number of anecdotal users who decided to raise this as an issue (as a result it's trending on X) says otherwise.
[0] https://x.com/thsottiaux/status/2090675027670978569
I love the fact that devs are still complaining that invoices are able to grow from $300 to $1000+
How can anyone use a platform where this is even an issue? Just because AWS is a failure in this regard, doesn't mean there aren't alternatives with fixed prices or others with easily settable limits.
Prompt edits leaking into the cache and affecting model responses is exactly the kind of billing-relevant behavior change that should be in release notes, not discovered by users.
Sounds like a path to profitability rather than a bug.
Bug for user = feature for company
Bug ... Feature
It would be ironic if this bug exists because it was vibe coded.
I wonder if it's related to Codex wearing out SSDs.
Rookie mistake - it seems like they didn't follow manufacturers' guidance when installing the 10x engineers. One needs to clearly define which metric should be 10x'd before powering them up.
Loaded question: would an openrouter or similar solution caught this before the $BigProblem showed up?
You know, installing unsloth studio lets you use codex against a local qwen 3.8 instance, which does ~10 tok/sec without GPU on a modern machine, and 100+ tok/sec on a 5090, and is incredibly good.
But but according to Mr Altman, we've like entered the singularity. Right? Who cares about a billing issue?
It’s called the singularity because all your money vanishes into a black hole.
Funny how it is always more charges but never less or no charges.
"Random" accidents that always go against you, too biased to be random.
But don't notice that too much, you might start to see patterns here and there that you're not allowed to, might get you banned from places, etc.
> Funny how it is always more charges but never less or no charges.
Recently: https://news.ycombinator.com/item?id=49337602
Quota resets are so frequent they don't make the news anymore.
I re-read that post/website several times.
Can you cite the portion of it that makes it clear that reduction happened by accident?
Applying Occam’s razor, which do you think is more likely:
1. OpenAI intentionally adds random overcharges. 2. OpenAI deprioritizes fixing actual bugs that cause occasional overcharges because doing so won’t affect their bottom line.
This can be a reporting bias. Noone opens an issue when they were billed too low.
I doubt anyone announces when they have under billed. OpenAI has also done many low price deals and quota resets.
Usually it's user's incentive to control over-billing and company's one to make sure there's no under-billing :)
But you have eyes, right? And a brain, and live through your own experience and can reason about it, right?
When was the last time someone undercharged you or didn't charge you at all by mistake?
Is this a common occurrence?
What's the proportion of overcharges vs. undercharges you have observed in your life?