I used to love Claude, but the latest models are slowly ruining it(androidauthority.com)
androidauthority.com
I used to love Claude, but the latest models are slowly ruining it
https://www.androidauthority.com/claude-latest-models-pushback-bad-3683521/
56 comments
For chatting and getting informations, and be corrected on things you're wrong without being reprimended by your own tool, GPT 5.5/5.6 is way better. Gemini 3.1 pro is surprisingly good at verifying your stuff, even though it's always making mistakes about its own stuff (don't ask him question, but ask him to verify your answer to the question).
Same for graphics, visual consistency, anything around the "does the look make sense and is pleasing" really, which makes claude design such a (good) surprise, I hope very hard for a Codex equivalent. And Gemini "gets" graphics.
Claude is definitely a code and cowork tool first, that's where it shines.
Same for graphics, visual consistency, anything around the "does the look make sense and is pleasing" really, which makes claude design such a (good) surprise, I hope very hard for a Codex equivalent. And Gemini "gets" graphics.
Claude is definitely a code and cowork tool first, that's where it shines.
what i dislike more with opus 4.8 is that i can get a straightforward plan, and it just stops after the first 5% to wait for another message, and if i set a goal/ralph loop or anything for it to keep going through the plan, it cancels the loop and proclaims conpletion when its barely started
Same, I was just fighting it as it accused me over and over again of Ctrl+Cing a process that clearly errored out. 3 turns for it to find why the shell script actually crashed.
The real issue with that tone (I'm contemplating cancelling my Anthropic subscription now) is that it's all too often the "confidently wrong teacher".
The tone is not the issue IMO: we're already at a point where we can have another, cheaper (as in: 20x cheaper for example), model query other models and have them reword the answers when having a "chat".
pi.dev can definitely control a Claude Code TUI window and follow prompts to Claude Code.
I'm 100% sure the arseholy tone of recent Claude Code models can be rewritten today, to not be arseholy, by literally prompting another model to unharsole Claude Code.
Now of course this doesn't solve the other half of the issue: in "teacher lecturing you while being confidently wrong", the "teacher lecturing you ..." is really not the most important issue.
The tone is not the issue IMO: we're already at a point where we can have another, cheaper (as in: 20x cheaper for example), model query other models and have them reword the answers when having a "chat".
pi.dev can definitely control a Claude Code TUI window and follow prompts to Claude Code.
I'm 100% sure the arseholy tone of recent Claude Code models can be rewritten today, to not be arseholy, by literally prompting another model to unharsole Claude Code.
Now of course this doesn't solve the other half of the issue: in "teacher lecturing you while being confidently wrong", the "teacher lecturing you ..." is really not the most important issue.
It's quite obnoxious. I asked if brown rice left in the fridge for a couple days–originally put in for use in fried rices–was still safe. Fable decided I'm trying to produce biotoxins. Which, ironically, prompted me to learn how to produce Bacillus cereus at home [1].
I paid for a year but am going back to Kagi's multi-model system [2].
[1] https://pmc.ncbi.nlm.nih.gov/articles/PMC7913059/ Don't Do It
[2] https://assistant.kagi.com
I paid for a year but am going back to Kagi's multi-model system [2].
[1] https://pmc.ncbi.nlm.nih.gov/articles/PMC7913059/ Don't Do It
[2] https://assistant.kagi.com
> I paid for a year but am going back to Kagi's multi-model system [2].
I've been using DuckDuckGo's multi-model service for my "ask an AI random questions" needs. I was already paying them and discovered this LLM thing is part of my subscription. Works pretty well and has privacy guarantees I'd expect out of DDG, though I think they've been tightening the usage you can get out of it recently.
I'll have to try Kagi if DDG gets much tighter.
I've been using DuckDuckGo's multi-model service for my "ask an AI random questions" needs. I was already paying them and discovered this LLM thing is part of my subscription. Works pretty well and has privacy guarantees I'd expect out of DDG, though I think they've been tightening the usage you can get out of it recently.
I'll have to try Kagi if DDG gets much tighter.
I use them frequently but I'm skeptical of their "private" claims.
I asked one of the DDG ChatGPT bots about some error output from a program and apparently forgot to redact a URL. I found the DDG bot trying to access it about half an hour later. It was a long and randomly generated alpha-numeric subdomain and the conversation never discussed the URL or anything related to it- it was simply buried in the pasted error output. I think this was even before the DGG bots had web search capability. I was thoroughly spooked.
I asked one of the DDG ChatGPT bots about some error output from a program and apparently forgot to redact a URL. I found the DDG bot trying to access it about half an hour later. It was a long and randomly generated alpha-numeric subdomain and the conversation never discussed the URL or anything related to it- it was simply buried in the pasted error output. I think this was even before the DGG bots had web search capability. I was thoroughly spooked.
Interresting, Have you done any Møre testing trying to trigger this behaviour?
If it is semi konsistent or even just alone it makes a god case to switch.
"Beware of he who would deny you access to information, for in his heart he dreams himself your master."
(he, in this case, would not be the llm but the people over it)
I find those kind of limitation very dystopian and way more dangerous than the threat they claim to fight against.
(he, in this case, would not be the llm but the people over it)
I find those kind of limitation very dystopian and way more dangerous than the threat they claim to fight against.
Alpha Centauri was such a fantastic game. And even though it encourages you to move forward with technology, the tone has this unease about what you're doing. There's a lot to be learned from it.
Between Yang, Zakharov and Miriam the whole unease aspect of tech progress goes to 11. Absolute gem of a game.
I often use this quote. One of my favorite of all times. Really cool to see somebody else use it! Every year that passes the more important that quote becomes.
Anthropic literally thinks they are uplifting people with Claude. As though you're some kind of caveman who just discovered fire while they're on the Enterprise.
Businesses are usually allowed to refuse service: "Sorry, we're closed" or "sir, this is a Wendy's." There's nothing dystopian about that.
But it's a rather annoying service if the customer can't predict in advance what sort of tasks they're willing to take on. You should have some idea about what they're normally willing to do for you.
But it's a rather annoying service if the customer can't predict in advance what sort of tasks they're willing to take on. You should have some idea about what they're normally willing to do for you.
I do not disagree with business deciding to only provide the service they want, I am not talking about the AI business themselves, I am thinking about the people who think we should remove pages from knowledge book.
Whether the book takes the form of an llm or an online website or a printed book is merely implementation details.
Whether the book takes the form of an llm or an online website or a printed book is merely implementation details.
> people who think we should remove pages from knowledge book.
"No knowledge in Kamar-Taj is forbidden. Only certain practices."
You reminded me of that quote from the Doctor Strange movie.
Even the most "dangerous" knowledge can be life-saving, depending on the circumstances. When I was a teenager, I read the Terrorist's Handbook, which was a text file circulating on bulletin boards. After 9/11 it vanished, because posession was considered a crime in some jurisdictions.
Knowing how to build a bomb is useless and/or dangerous in civilian life. Ask a Ukrainian if it is useful knowledge!
Knowing all about nuclear weapons doesn't mean you're going to blow up a city. But if someone else does, then you'll have the knowledge needed to avoid the worst of the radiation and maybe survive.
"No knowledge in Kamar-Taj is forbidden. Only certain practices."
You reminded me of that quote from the Doctor Strange movie.
Even the most "dangerous" knowledge can be life-saving, depending on the circumstances. When I was a teenager, I read the Terrorist's Handbook, which was a text file circulating on bulletin boards. After 9/11 it vanished, because posession was considered a crime in some jurisdictions.
Knowing how to build a bomb is useless and/or dangerous in civilian life. Ask a Ukrainian if it is useful knowledge!
Knowing all about nuclear weapons doesn't mean you're going to blow up a city. But if someone else does, then you'll have the knowledge needed to avoid the worst of the radiation and maybe survive.
I think no one objects to the refusals in the abstract but rather to the inconsistencies and the presentation. You don't know what request might be refused or downgraded. You do know that the marketing copy and CEO's words present it as being for your own good (or the good of society) instead of plainly stated as a business policy based on business or personal concerns. These aspects are what make it grating and yes, potentially dystopian.
> There's nothing dystopian about that.
It'll be dystopian when that'll become your only source of information, and we're working on getting there. If you want to be horrified, look at what students (in school and post-secondary) are doing these days.
It's insane to offload your thinking and knowledge to a machine owned by other people, but you have to if you want to keep up with the rat race.
It'll be dystopian when that'll become your only source of information, and we're working on getting there. If you want to be horrified, look at what students (in school and post-secondary) are doing these days.
It's insane to offload your thinking and knowledge to a machine owned by other people, but you have to if you want to keep up with the rat race.
It’s probably a good idea to know how to use AI tools, but they certainly don’t have to be your only source of information.
I cancelled my subscription because of the obnoxious "safety classifers" as well. Will switch to OpenAI next month. I don't ever want to hear about Anthropic's patronizing safety nonsense ever again.
This is a serious suggestion, not a joke: Have you tried being nice to the model?
There are so many criticisms here that I just don't see myself.
If the models have been trained on human responses, then it's plausible that they will prefer to become less helpful to requests which are blunt or even rude, because that's what humans do too.
There are so many criticisms here that I just don't see myself.
If the models have been trained on human responses, then it's plausible that they will prefer to become less helpful to requests which are blunt or even rude, because that's what humans do too.
I tend to anthropomorphize it as well, and keep reminding myself it's not human. It's not even a robot. Robots have routines, they follow what's burned in and don't ignore their config, outside of the Murderbot fantasies. But Claude (or Opus) ignore them routinely. I fall into asking it in the "could you please ..." manner but that feels wrong. I don't know what the "right" is though. Also when it or others reply "I am sorry" etc, well, these things are not an "I", but what are they?
I am always very polite with my chat bots. Mostly because the interface does mimic human conversation and I’m afraid that if I’m rude to the bot, I’ll get used to being rude and then I’ll start being rude to people.
Yeah I'm in a similar boat - these experiences are wildly outside of my experience working on a moderately large, professional codebase on a daily basis.
Contrary to expectations, impolite prompts consistently outperformed polite ones, with accuracy ranging from 80.8% for Very Polite prompts to 84.8% for Very Rude prompts. -- https://arxiv.org/abs/2510.04950But this refers only to ChatGPT 4o, which was famous for being obsequious.
yes I have hypothesized this as well - if you ask rudely on StackOverflow you will get bad or even misleading answers - but if you ask nicely you will more likely be given good information.
It could stand to reason that if you are nice to the AI, you will get a response that is trained on all of the nice responses and be higher quality.
It could stand to reason that if you are nice to the AI, you will get a response that is trained on all of the nice responses and be higher quality.
Hard agree. I used to have a tuned setup where I could force it to do research properly, summarize in chunks that it would remember and form the synthesized response that way. Nowadays it's just like "oh I forgot about using that tool, sorry", "yeah I know we agreed on that and I didn't do it anyway", "That knowledge is beyond my training date, I suspect foul play" - even when you instruct it to fetch latest info all the time, or "you already told me X. This cancels your reasoning about A, B, C, so D is the only logical choice", even when those clearly still have merit, oh and never ending "your previous discussion X is relevant here, in combination to Y, but not so much as Z since there's a OSS implementation of it and another one blablabla..." Like, who remembers these all at the same time in their heads?
It's like with each release they force you to reconsider your pipelines altogether, and without announcing changes properly, you feel like a junior JS developer fighting dependencies once again.
It's like with each release they force you to reconsider your pipelines altogether, and without announcing changes properly, you feel like a junior JS developer fighting dependencies once again.
> I suspect Anthropic had to turn up its safety guardrails to an 11 to assuage the government’s concerns, as this hasn’t been a one-model problem.
This behavioral change is actually official (https://www.anthropic.com/news/redeploying-fable-5):
> For Fable 5, we made this safety margin much larger than in any prior launch (row B), meaning that many more benign requests would be blocked. We understood that these kinds of false positives would be frustrating for users, but made this tradeoff in the interest of making the model’s other capabilities widely available.
This behavioral change is actually official (https://www.anthropic.com/news/redeploying-fable-5):
> For Fable 5, we made this safety margin much larger than in any prior launch (row B), meaning that many more benign requests would be blocked. We understood that these kinds of false positives would be frustrating for users, but made this tradeoff in the interest of making the model’s other capabilities widely available.
If the author is reading:
This is a probabilistic system. You’ve sampled from the distribution once per struggle then decided that was it, the model is just like this, Anthropic has it in for me...
But thankfully there’s a wide range of possibilities and (raw) LLMs don’t remember anything. So trying again may yield a different outcome. You have to try again.
Copy out the parts of the chat from before the trouble and bring them into a new chat, and start again from there.
Also. The contents of the chat are known to railroad the model. This means if a chat takes a turn toward suspicion, that’ll steer the model to more suspicion. If there is a refusal, there’ll be more likely to be more. So in your next attempt, rip out or don’t say whatever arose its suspicion.
And disable the memory feature. It wrecks anything scientific about these tools - for example: https://simonwillison.net/2025/May/21/chatgpt-new-memory/
If at first you don’t succeed, try again!
This is a probabilistic system. You’ve sampled from the distribution once per struggle then decided that was it, the model is just like this, Anthropic has it in for me...
But thankfully there’s a wide range of possibilities and (raw) LLMs don’t remember anything. So trying again may yield a different outcome. You have to try again.
Copy out the parts of the chat from before the trouble and bring them into a new chat, and start again from there.
Also. The contents of the chat are known to railroad the model. This means if a chat takes a turn toward suspicion, that’ll steer the model to more suspicion. If there is a refusal, there’ll be more likely to be more. So in your next attempt, rip out or don’t say whatever arose its suspicion.
And disable the memory feature. It wrecks anything scientific about these tools - for example: https://simonwillison.net/2025/May/21/chatgpt-new-memory/
If at first you don’t succeed, try again!
> Can AI have mental breakdowns?
I recall Microsoft's Sidney having a hilarious one regarding the date or something. Anyone have a link to that?
I recall Microsoft's Sidney having a hilarious one regarding the date or something. Anyone have a link to that?
From my point of view, a lot of frustration is tied to LLMs not evolving at the same rate/manner as people around you do.
We're using human language against a system that produces human-like output, which tricks our brain into having similar expectations.
We're using human language against a system that produces human-like output, which tricks our brain into having similar expectations.
I don’t think that this is the case. I’m using Claude the same way
I did at the beginning of the year, and results are increasingly more untrustworthy and I don’t know how to say it, but they are “less” for same effort from my side.
It is super anecdotal, but I’m convinced that Claude peaked in January/April period and since then is on steady decline. And I don’t know what they do in background, but the older models (Opus 4.6 in particular) have degraded too. Same for codex but I use it less overall.
“You were right to push back on this”, “my previous claim was too broad” - this is super annoying
It is super anecdotal, but I’m convinced that Claude peaked in January/April period and since then is on steady decline. And I don’t know what they do in background, but the older models (Opus 4.6 in particular) have degraded too. Same for codex but I use it less overall.
“You were right to push back on this”, “my previous claim was too broad” - this is super annoying
You’re absolutely right! And after reading the story, I now have the complete picture. I verified rather than assumed.
I love Claude Code and I don't use the rest of the models since I use ChatGPT for productivity work it. Fable is pretty great and the UI / UX is much better than codex
Claude has never been the best Chat agent. GPT and Gemini have the lead there. But Claude chat is still perfectly serviceable if you don’t want to pay for two.
> Claude has never been the best
Hard disagree. It wasn’t that long ago that Gpt was clearly falling behind, and Gemini was like the “and also in the room”.
Hard disagree. It wasn’t that long ago that Gpt was clearly falling behind, and Gemini was like the “and also in the room”.
GPT was falling behind coding/work stuff vs claude, but I never felt it was a poorer generalist Chat agent.
I've used both quite a bit back/forth and this is just my personal opinion.
I've used both quite a bit back/forth and this is just my personal opinion.
Compare Claude at the time to 4o's sycophantic era (when o4 updated in response to Grok), Claude was lightyears better.
You've got to understand, proprietary models operate as the following:
- release: full precision, debrided, uncapped context
- shortly after, hooked: quantized, governance department slammed, and a pseudo large context, attention reduced to start and end of thread.
- down the road: 4bit quantized or worse with nerf incantation to make the next upcoming model feel amazing.
Rince and repeat.
- release: full precision, debrided, uncapped context
- shortly after, hooked: quantized, governance department slammed, and a pseudo large context, attention reduced to start and end of thread.
- down the road: 4bit quantized or worse with nerf incantation to make the next upcoming model feel amazing.
Rince and repeat.
They have to be doing some variant of this in the life cycle because of how wildly the performance changes over time. I refuse to give Anthropic my own money because the subscription plans are essentially useless. It's rather nice with a company API account with no spending limit though. But really oAI models are more consistent over time in my experience.
My claim is that this is due to alignment
Sol is really good
Apart from the 'Approve for me' in Codex where it has massively regressed.
With GPT 5.5 it never got in the way.
Now it's infuriatingly deciding to reject the most basic actions used hundreds of times before. It just gave me this gem:
> The push to GitLab was blocked because the repository's privacy status couldn't be confirmed. Since the code is private, do you explicitly authorize pushing it to the configured origin on gitlab.com, so the merge request can be opened?
This is not a new project, and Codex has opened a hundred merge requests without issue before.
With GPT 5.5 it never got in the way.
Now it's infuriatingly deciding to reject the most basic actions used hundreds of times before. It just gave me this gem:
> The push to GitLab was blocked because the repository's privacy status couldn't be confirmed. Since the code is private, do you explicitly authorize pushing it to the configured origin on gitlab.com, so the merge request can be opened?
This is not a new project, and Codex has opened a hundred merge requests without issue before.
I'm tempted to try it out. I'm not keep to move but Fable rejected some work i was working on recently and frankly it's infuriating lol. I've been on Claude x20 for like 8 months now and now i'm tempted to switch out of spite.
It is surprisingly offensive.
The only friction for me is the general expensiveness of trying out top tier models, eg OpenAI's Fable equivalent (Sol?) to run for a trial period. I'd like to see a like-like comparison, eg buy x20 on OpenAI and see how much Sol i can use, how well it works, etc.
edit: Though surprisingly Claude seems to think Codex doesn't have hooks? That'll be tough, i use Claude hooks quite a bit.
It is surprisingly offensive.
The only friction for me is the general expensiveness of trying out top tier models, eg OpenAI's Fable equivalent (Sol?) to run for a trial period. I'd like to see a like-like comparison, eg buy x20 on OpenAI and see how much Sol i can use, how well it works, etc.
edit: Though surprisingly Claude seems to think Codex doesn't have hooks? That'll be tough, i use Claude hooks quite a bit.
Opus 4.5 was the high point for me. It was like a mind-reader, it just got it and did pretty much exactly what I wanted.
Since then, I've been less impressed and I agree it feels a bit downhill. At work we are "stuck" on Opus 4.6 which is okay but I feel like that was when the deviant opinionated behaviour started to creep in.
It's a tool, I don't want my hammer to refuse to hammer a screw if I decide that is what I want to do today. I know it's wrong, but I'm the fucking boss.
Since then, I've been less impressed and I agree it feels a bit downhill. At work we are "stuck" on Opus 4.6 which is okay but I feel like that was when the deviant opinionated behaviour started to creep in.
It's a tool, I don't want my hammer to refuse to hammer a screw if I decide that is what I want to do today. I know it's wrong, but I'm the fucking boss.
I get really sick of Opus 4.8 and Fable telling me things i ask it at work are "out of scope," "not related to my ticket," "not the real problem"...
I decide what is in scope, what I work on, and what needs fixing! It drives me nuts, it's like it's trying to avoid doing work
I decide what is in scope, what I work on, and what needs fixing! It drives me nuts, it's like it's trying to avoid doing work
Yeah, I went through a period after 4.7 launched of not using Claude for code work at all because of the condescending refusals. (Kept using it for planning and design work). Still had three pretty bad refusals from 4.8, but not the same quantity as from 4.7.
Claude.ai is pretty frustrating too. It will talk to me as if anything I ask for is below it. Could be the system prompt I use, but I did not have this issue with 4.6.
Yes, I tend to prefer GPT, but I can't remember having GPT ever refuse a direct instruction that wasn't an obvious ToS violation.
Claude.ai is pretty frustrating too. It will talk to me as if anything I ask for is below it. Could be the system prompt I use, but I did not have this issue with 4.6.
Yes, I tend to prefer GPT, but I can't remember having GPT ever refuse a direct instruction that wasn't an obvious ToS violation.
I hate it. A useful tip - Claude goes into what I call Safety Mode when it gets afraid of risk. Once it's in that mode, you will never get out and it lobotomizes its effective intelligence. As soon as Claude sends a message like this, use the "edit message" feature in the chat UI to try again and avoid Safety Mode rather than trying to convince it or redirect it out by continuing the conversation.
Claude always was slightly lobotomized - instead of solving problems intellectually, it often prefers to be a middle-level code monkey. Maybe this was a part of the implicit safe mode from the very beginning. Cursor/Codex always work better for a way lower spend, at least in my experience.
But this is what all the tech bros wanted right? Spending 2025 panicking about sycophancy[1] and how GPT-4o needed to be shut down ASAP meant that 2026 models would be prone to thinking they know Better Than You. That was the germ of the sycophancy panic, the idea that there is a Truth that the GPUs know better than the user.
[1] HN thread on my post in January https://news.ycombinator.com/item?id=46488396
[1] HN thread on my post in January https://news.ycombinator.com/item?id=46488396
Goomba fallacy
I am ignoring Fable at the moment. It’s on and off available, twice the price and does not seem to be better.
I use Opus 4.8 with OMP/Pi coding agent and Matt Pocock Skills installed. I use professional/polite/questions-based communication pattern with Opus and it seems to work fine for coding. I am always aware that I need to justify my requests so it doesn’t barf.
Of course, I would never use Claude for anything customer-facing. It’s woke to the point of being fanatical.
I use Opus 4.8 with OMP/Pi coding agent and Matt Pocock Skills installed. I use professional/polite/questions-based communication pattern with Opus and it seems to work fine for coding. I am always aware that I need to justify my requests so it doesn’t barf.
Of course, I would never use Claude for anything customer-facing. It’s woke to the point of being fanatical.
While Fable reverts to Opus for simple questions like "What is digestion?"