Any benchmark showing Opus 5 as the best just loses credibility for me. Anyone who's actually used Opus 5 daily knows what I'm talking about.
With weaker models you can sort of understand, they're trying their best and failing, but this thing just channels its immense inteligence into being as annoying as possible instead. I know it can do what I'm asking it to do, but it just finds a way to weasel out of it, or maybe just thinks for 10 minutes instead, then fixes one thing and breaks four additional ones.
But even fable has the annoying tendency to invent new jargon and produce an incomprehensible soup of text.
"give me this again without jargon invented this session at high density
and with a couple (maybe more or less) simple useful ascii diagrams underneath each design"
The context is that I was discussing an experimental new idea for my video game review analysis product.
Designs 1,2, and 3 were horrible: the model even suggested a rejection after the word soup so it would have been pointless to waste my fleeting time on Earth reading it.
Otherwise, I generally really enjoyed using fable for bouncing ideas. It was an absolute joy to have this thing provide useful criticism, analyse sample data, and create prototypes so that I could elevate my understanding of the problem without stepping down from a pure intuition/design headspace.
But I don't consider the purely model written code usable for a feature this important. I'll probably scrap it entirely and start from scratch with newfound understanding.
> Monitor clipping. review_watched produces PathStatus and nothing else. It never feeds solve. Whatever it does to legs cannot reach the search.
(that's after being told twice to not use shorthand jargon nor reference the code directly)
It's a simple switch to make: cursing = try harder instead of cursing = stop trying. Is it really impossible to train Claude that way?
Especially the second one seems exactly like my experience.
After I started reading complaints about Opus 5, I gave Fable the task of evaluating a bunch of code Opus 4.8 had written and compare it to Opus 5's code. Fable ran a dynamic workflow and the scores came back 15-20% higher for Opus 5's code in terms of quality, correctness and readability/conciseness. I did not tell Fable which Opus wrote which code, and I turned off memory as well to ensure there was no pollution from that angle.
My only complaint is that Opus 5's prose is annoying as hell. I wrote a custom skill for it for concise debriefs and it has been working pretty well for me.
I’m doing the same right now, and I’ve found that asking for “simple English” works most of the times, although not always.
Did you find better wording that works consistently?
I'd open a blog with "weird things Opus did". Today it launched a swarm of cpu-hogging processes to test if the widget showing machine and I/O load is rendering nicely and correctly. The test went fine, but it was no longer able to kill those processes since they were really effectively hogging the CPU in various ways - being diligent, some of them were hogging CPU, some were murdering the SSD, some were pounding on the network adapters. Took me 30 mins to recover the machine to a working state without killing the meaningful, messy, in-flight sessions i had going on on other projects.
Infuriatingly so, in a way I don't remember Opus 4.8 being, but maybe I've just been ruined by Fable 5.
I got so used to it, when they finally pulled access for me and I had to go back to Opus I felt like I was working with my hands tied.
I finally know what those women with AI boyfriends felt like when their app updated and it won't dirty talk with them anymore.
Like the Fable ban stunt, I wouldn't put it pass Anthropic to kneecap Opus deliberately to drive more people to their more expensive option.
To me it's not so much the dumb mistakes (although there are some of those) but the ultra-verbose, mega-inefficient "solutions" to some problems / prompts.
Stuff that "works" if you're the kind of person that considers slamming a semi-trailer at 200 mph into a door did, technically, result in the door being somehow "open".
As it's supposed to be one of the most advanced model, I can't help but wonder if the solutions are that bad/verbose/inefficient because we're already in a loop of models being trained on sloppy-pasta from previous models.
“One thing worth your attention, if you were to detonate a pipe bomb in your house, it would have a negative effect on your living room”.
It will do everything it can to defer or push it off, to the point where I’ve had to add multiple imperative directives to the AGENTS file telling it, in no uncertain terms, not to defer tasks under any circumstances.
• Qwen3-VL picks up new images in a NAS, auto captions and adds the text descriptions as a hidden EXIF layer into the image, which is used for fast search and organization in conjunction with a Qdrant vector database.
• Gemma3:27b is used for personal translation work (mostly English and Chinese).
• Some small 8b models (like llama3.1) for sentiment analysis on text.
But haven't really tried using local LLMs in conjunction with agentic harnesses yet.
My secret sauce is to use LLAMAcpp's reasoning-budget and reasoning-message that trigger cut off to overthinking with a message that says to either us subagents or compress the context. opencode's dynamic context pruning plugin can get you pretty far into the stratosphere.
Thanks for the tip - I like this a lot. I remember having to do a lot of tweaking to curtail Qwen QwQ-32b when it would go down an endless psychotic recursive reasoning loops as part of its "chain of reasoning."
I havent yet tailored the pruning messages, but mostly it works.
Reasoning budget can also be set by client, so potentially smarter.
See https://news.ycombinator.com/item?id=48883538 25 days ago
> The Sleev (the project has been renamed to make a startup) creator was shilling their project in the OpenCode Discord. That person is very convinced they have something that no one has ever built before. They focused on token reduction without any real evals for capability impacts.
I'm generally against this context pruning without prompting or details. Sleev is very opaque about how it works and definitely will bust your cache.
There's a few times it gets too dumb to cope, but in comparison with other coding agents, it seems smarter.
Opus 4.6 is the last model that's actually useful and can "adjust" its perspective to use the newer & better solution.
Where Opus 4.8-5 has over fit training on worse/older but "dominant" solutions it refuses to adjust.
Not only does this create an existential threat to adopting progress but it also means that if you have a code base that has rare but real world tradeoff the newest versions of Opus 4.7, 4.8 and 5 are worse than useless and become a major dev timesink.
I could not get Opus 5 to do anything without losing a few years of my life from stress.
Fable has been okay but I am doing ML work and not allowed to use it which feels insane.
I don’t see anyone talking about how you have to completely change your prompting strategies with Op. 5 versus 4.8 to get the most success.
Note this is for complicated tasks - simple stuff can be done by whatever pretty well.
Also models will be awesome at doing something at 150k tok of context and terrible at 750k toks so even within the model itself there are capability considerations.
It’s an engineering problem to design around, not something you can escape.