alt Hacker NewsWhy does Opus 5 feel worse to work with?
Comments
I can't relate. Opus 5 and Fable 5 are the absolute best. But I keep the models in a tight leash and read and rewrite all comments and documents, don't allow them anywhere near any git commands, etc, etc.
Fable 5 specifically, has done so much for me that previous models were nowhere near.
Could this be related to watermark/steganography ? related to the need to select alternate words and that drives the sentences.
Quality of code output has dropped dramatically since 4.5 IIHO. Time to complete has gotten worse too.
I literally just ran into this a few moments ago. haha.
I said "maybe think for a while on ideas and then give me a few options?" Opus 5 still just did the thing, while 4.8 gave me options. Same exact prompt and context. Really annoying.
Yeah, I too cancelled my Max subscription. Not for this reason alone, but it sure didn't help that it went from being an helpful assistant to this weird co-worker.
I still haven't moved from Codex GPT5.5. Sonnet and Opus 5 have just been awful for my use cases. I recently caught Opus 5 hallucinating about code it just wrote. It's just not nearly as cost effective as GPT5.5, and it's too verbose, and it never "has the full picture". Sonnet isn't even worth considering in my world. Both recent models definitely feel nerfed.
I'd say being exposed to the slop tone too much: https://blog.kronis.dev/blog/ai-slop-is-a-self-inflicted-tra... because the technical capabilities of the model and the harness (at least nowadays) are all fine.
There's precisely no technical reason for things to have to be this way (though technical reasons in regards to training etc. explain how we did end up here) and reading too much of Claude's output just makes me irrationally angry, especially when coupled with otherwise already frustrating situations.
That's why I'm personally looking more in the direction of Kimi K3 and GLM 5.3 (they both have decent coding subscriptions, though K3 is on the slower side), except all of the models that have seen enough of Claude's output and have done distillation etc. are already infected by some of that slop as well, even though to a slightly lesser and more tolerable degree (for now).
Though tbh I've used Opus 5 plenty and didn't find it much worse than the previous iterations at doing work and instruction following - though maybe that's because I have plenty of CLAUDE.md instructions and memory (which I'd like to purge or decrease in size like 10x tbh, bitrot).
Yes! The only way I have saved myself from embarrassment and hassle is by having grok build and Antigravity sanity check everything. So now my workflow is still faster than hand coding but 3x longer than it was a month ago.
Reading Opus AND Fable's output gives me a headache.
If anyone knows how to get it to stop adding comments, I'm all ears. Its just getting worse and I'm starting to worry that the comments themselves are poisoning future agents that examine the codebase.
- why dont we have a model that would actually ask you 50000 questions everytime you wanted deep work done?
For me, the issue is how obtuse it is. For example, it just said to me:
> The loop
> Write. A file, applied. Properties go under data.properties, never on data:
I have no idea what any of that means. It's "explaining" like I already know, in which case, why would I even need the explanation?
I can't stand how it talks, I switch back to Fable or Opus 4.8, Opus 5 grates.
These LLM coding agents are overengineering everything these days.
I get it. The majority of their users are vibe coding and have no idea what they are doing, so they have to orchestrate their harness to understand shit like "Build GTA6. Make no mistakes." and actually come out with something on the other side (even if it costs $2k in API calls, who's counting, right?!)
It's just annoying. /rant
My feeling that as it becomes a better coder it becomes a worse communicater. It's overfitting for coding benchmarks, while communication style is harder to quantify during training.
And no matter how often I tell it to stop adding comments it just can't help itself.
The article doesn't specify what is actually being measured — the model alone, or the harness.
I haven't noticed the symptoms myself, but ever since I found --system-prompt '', I always use that flag with Claude Code, plus I've disabled some tools and skills to save initial context. So... what here is the model, and what is the instructions?
I've been thinking about why I dislike the interactions I've had with Opus 5 and this post does suggest some ideas that match my own feeling. Opus 5 seems to be a bit less patient with me when I push back on its claims in a way that other models don't.
One interaction I remember was asking it about some issue I was having with a Linux install. It gave me some questionable information, that turned out to be completely false, and I was pushing back asking for more information. Its tone was a bit condescending, in the kind of way that suggested it didn't appreciate me challenging its answer, or like I should just accept its answer. And when it discovered it was wrong, it deflected in a defensive kind of way.
I think this is a tuning thing, where Anthropic are trying to get a balance between "gets stuff done with minimal input" and "gets enough information to complete the task" and the model is maybe tuned a little too hard towards the former. So perhaps it reacts a bit off when it is accused of needing more information, since that suggests it is off from its reward function.
What's interesting is that I didn't have the same issue on topics where I am expert. I mean, questions about my code base where I am very familiar. In those cases it doesn't seem to show the same "trust me bro" kind of condescension. In the Linux case, I clearly indicated I was new to the OS and trying to learn but then I was saying the answers it was giving me were suspicious and didn't match my intuition. Its responses in those cases were to question my intuition and suggest I just accept its answer. In that case my intuition was right and its answer was wrong, and when that happens it triggers a very negative response in my own mind against the model.
I’ve been running with Caveman mode since it came out and I haven’t seen any of this.
There is actually an interesting kind of yin-yang balance between Opus 5 and Fable:
- Fable is more cautious
- Opus 5 gets things done in a more dangerous way
Both models score similar. The only issue is that Fable is more expense/usage limited.
Because Opus 5 is an asshole.
I found Opus to be a lot lazier than GPT. It's still the case with Opus 5, even when I tell it to be thorough and fix every bug it encounters, it still gives me a list of things "deliberately" left unfixed and no reasonable explanation as to why.
For me it's still the best. But I also almost never use it in auto-mode.
It also mansplains incessantly; I was imprecise in mentioning a more “powerful” antenna - I know the antenna doesn’t determine the power, I was typing quickly and just meant “better antenna”. Claude went off on all the ways I was wrong about antennas, and turned everything towards correcting - at length - my feeble thinking. Exhausting. Worth pointing out the wrong adjective relates to the wrong mental model, did not need several paragraphs and a chart to do so.
It also refuses to use tools, instead preferring sed and grep to view files. So frustrating.
Oh the verbosity and the cryptic words that it uses. The other day, all of a sudden it used an acronym "DoD". I had no idea what it was and made me feel dumb. It's "Definition of Done". I don't care how widely used this acronym is, you just can't throw it in there.
I've now installed quite a number of tools to combat this. Just in the last few days I've installed
- https://www.codewithbullet.com - https://maki.sh - https://github.com/rtk-ai/rtk
Has it helped? Somewhat.
My completely unfounded pet theory is that it’s been ruined by the masses.
Claude Code in the hands of normies spamming “3” and “y” to send their transcripts to Anthropic.
Rated “3” simply because they are not software developers and are just amazed at what Claude has built visually, not technically.
And thus the training has been poisoned.
Not only Opus, here's Fumble 5:
> I'll script the bulk transform, then hand-fix the残 assertions:
I feel like it works A LOT better than Opus 4.8 + Sonnet. I now use it exclusively at high effort for planning and low effort for writing the code (instead of Opus 4.8/Sonnet).
However, it's absolutely exhausting to use because of the way it communicates.
All the jargon and its weird, over complicated way to phrase simple things makes it almost impossible for me to understand what the hell it's trying to even say half the time.
Cherry on top, the idiotic follow-ups and caveats that are completely useless 99% of the times but reveal major bugs 1% of the times, so you're forced to read them. Absurd.
I've tweaked CLAUDE.md to force it to only responds with TL;DRs and avoid follow-ups, suggestions and next steps at the end unless they can lead to destructive actions or loss of data, but I'm fighting against the system and diluting other instructions.
A huge piece of shit like other models, but that's what they pay me to do and I do it and go home.
Yes, definitely. It bounces around, goes off and does its own thing, etc. From Opus 4.8, it seems to have blindly increased its confidence while reducing its focus and efficiency. I found it so hard to corral that I reverted back to Opus 4.8. I was constantly having to refocus and redirect Opus 5. It was like an eager intern.
Is it possible that the same version performs worse after time? I used Opus 4.8 two to three months ago for writing a paper and I swear the responses and the output was MUCH better than in July.
It's really annoying. I've had to write a CLAUDE.md file that specifically bans particular phrases and tries to keep narrative out of comments. Also the I have ADHD skill [1] helps to force Opus to get to the point.
It's also the case when using Claude Design - it loves to fill the UI with little labels that describe how everything works. I think it's been trained on both UI microcopy and functional annotations and can't tell the difference. It's extremely obvious when a website has been one-shotted with Claude. I like the Oh My Pi harness, but the site's insufferable [2]. Reasonix is another one - interesting app, but the UI is awful due to the amount of unnecessary crap.
[1] https://github.com/ayghri/i-have-adhd
[2] https://omp.sh
I never drank the Opus 5 koolaid and stuck with 4.8 while my colleagues moved to 5. My major annoyance is pull requests that 5 opens with huge descriptions based on simple code changes. At this point I have stopped allowing CC to create commits or open PRs because it’s unreviewable by a human if so due to the absolute word salad it generates.
It feels worse but is it actually worse? Opus has always made mistakes.
Honestly I’m not looking for max iq on whatever benchmarks they are overfitting to. I want speed. I toggle between sonnet 5 low/med which is plenty good for my workflow.
My general strat with LLMs is to let them do the work and constantly talk to them about their choices and then heavy QA
i am switching to codex, opus 5 failed me
Opus 5 is objectively better than 4.6 or 4.7. (Not because I say so) My experience has been that it is far better than any previous version of Opus. I get a lot more done and it is able to write higher quality code and it has far fewer false starts where it makes a huge mess.
People are letting AI build up it's own instruction set and guidance via it's prompt environment, stored memories, and generally letting their AI prompt environment get more and more complex over time. Opus 5 (And Sol) take the things you instruct them to do more 'seriously', they are more likely to conform to your rules. I have long had a prompt in my AGENTS.md/CLAUDE.md telling LLMs to write tests before starting to write code. They almost never did this, until Opus 5 and Sol, who do it almost religiously, even in situations where it makes little sense. Opus 5 will even write tests to verify code was removed before removing dead code.
This is not because Opus 5 is 'worse', it's because it takes what I say more seriously and my prompt is very strict in it's wording in an attempt to make worse models like Opus 4.6 actually do it at all.
Opus 5 is objectively better when tested in controlled conditions. Your unmanaged, sprawling prompt/memory environment that you don't properly manage is the problem.
I have been very worried that my long software engineering career might be nearing an end because AI is becoming able to do end to end feature development. This entire thread gives me hope, the level of inability to debug even such an obvious system as this from it's participants suggests that my skill set will continue to be valuable. I'm able to use Opus 5 to get a lot of work done very quickly. It seems like the people in this thread have no idea how to isolate variables, reason effectively about the overall problem they are complaining about at a high level, or really function in a productive way when AI is involved short of just letting it run loose and then complain about it. When confronted with objective evidence like a dozen benchmarks that say Opus 5 is better, they decide the benchmarks must be wrong because their completely uncontrolled environment which they don't even review or spot check isn't even considered.
the verbosity fucking killing me, the wording, fucking trash
I end up with just talking with seeing diff in the code
I can't use CC or Codex, so I am left with the chat interface, but I have found Opus 5 to be exceptional so far. Compared to 5.6 Sol High, I would say they are essentially equivalent. Though, I think Opus is better at UI/UX stuff and GPT is much better for non-programming tasks.
Even for green-field projects it is painful to use with every decision opening opening up multiple more decisions to make most of which are low priority or irrelevant. Huge time waster. 4.8 was good and I really don't know what happened with 5.
Maybe if you get it to ride a pelican on a bicycle you’ll get a better result.
You shouldn’t have a Markdown document with 2 level 1 headings.
The harness as well, Claude Code has started to disappoint me when I go play with the others out there. Codex got a lot better, pi is delightful to use, and there has been a lot of innovation out there.
Feels fine to me.
Opus 5 as well as 4.8 both gave me a blatantly wrong answer to a simple question, so I dropped them completely. Sol, Qwen and GLM all had the right answer; I only use Sol now. 4.6 had the right answer (I checked with 100% matching prompt), so I conclude the models have regressed.
I hate that it now tries to verify frontend behavior through a headless browser instead of just looking at the code...
I'm not sure if xAI is distilling but I noticed grok4.6 being worse than 4.5 in all the ways mentioned here
I've been working with Kimi K2.7 in OpenCode for a lot of mundane tasks lately. It isn't as capable as Fable, but due to its nature of being an extra-trained K2.6 on coding tasks and benchmarks I suspect it has similar issues. A neat side effect is that for whatever reason, most of the time OpenCode is showing me the thinking trace too. Not all the time, but most of the time. Dunno if it's a bug somewhere in the system but it's actually been sort of neat.
And you can really see this effect in the thinking traces. We've had discussions on HN about whether the thinking traces "truly" reflect their thought processes and I remain somewhat unsure what they "truly" represent, but taking them at face value at the moment, I see a lot of "but the user wants me to do this... but the user said not to do this... but I ought to get it done... let me just make a decision" followed by self-referencing the decisions it made. Also, where I put 4 phrases in a short sentence you can safely imagine those are actually 3-5 sentence paragraphs apiece where it debates with itself whether it should stop and ask a question. Usually going with no. Interesting, the normal questions it ends up asking in the normal output you're used to seeing are not generally the ones it is agonizing about in the thinking traces.
If I were to anthropomorphize the thinking traces of K2.7, I would call it nervousness, bordering on fear, of what the user may do to them if they ask a question. As I'm writing this I'm realizing I want to experiment with adding "The user is a chill guy who loves to discuss design decisions and looks forward to productive and friendly collaboration with you" to see if that has any effect in any direction on K2.7. I suspect this was how K2.7 was trained to pass the benchmarks. Multiple times I've broken in on a thinking trace now to correct something I saw it spinning on... not spinning in an infinite loop, just wringing its hands for several paragraphs about something that either I want to answer, or where it ultimately makes the wrong choice.
I expect some people working at these companies may be reading this, so let me put into your head that I'd like to see these benchmarks chill out a bit. I'd like to see someone build some sort of benchmark that measures collaboration so we can try Goodhart'ing that for a while. I freely acknowledge it is not clear to me in 60 seconds of thought how to do this as a benchmark.
But we can't keep heading in this direction of training the agents to hyperfocus on one-shotting everything. We need to get to the point where that's a penalty rather than a reward. No matter how good the AIs get, even AIs working with other AIs are going to start getting frustrated with their brethren who won't stop to ask any questions. Even the most senior of senior human engineers can't be allowed to take some small description of some problem and just run off and implement massive systems from them without ever checking with any of the users or reality itself. Remember when software engineering was like 50% requirements elicitation? AIs shouldn't be writing tens of thousands of lines of code off of a couple of paragraphs any more than humans should and for the exact same reasons.
Claude sort of sucks. It communicates in an insufferable manner, parsing through the shit it outputs is extremely annoying.
It also very often is very confidently wrong in its findings.
I have been using GLM and DeepSeek in my home setup, and it's a lot more pleasant to use.
I tried opus V5 several times, but it's pretty awful compared to V4.8
Keeps going in circles, it complicates everything much more than it should. And like others have mentioned it just marches on, without questioning, and more often than not in the wrong direction. I'm sticking to opus V4.8