It's not just you, Opus 5 is a "Jargon Douche" - but there's a fix

Tl;DR - There's been a very strong degradation in the default language outputs of Opus 5. It's as though it's using english, but the english is weird. There's a fix!

It's not just you! Opus 5 is a "jargon douche"

Here's the thing, like with the GlazingGate of last year, these types of "issues" are hard to notice on a personal level. We don't often chat with Claude in front of others, and we don't often discuss the outputs of our chats, so noticing something like this requires.

So often, this requires someone else to share their experience, for you to say "oh, I've noticed this too, I thought it was just me!"

I first noticed this when testing Opus 5 over the weekend post release, but didn't really think about it much. After all, it was used to create full 3d games using @mattshumer_ Gauntlet prompt, and folks were focusing on performance and evaluations most of all.

But then, the timeline and the evidence started to add up, folks started sharing their experiences with Opus 5 (and to some extend Fable) and more and more added to the realization, that it's not just me, Opus 5 really is a bad communicator.

Then, thanks to @yampeleg and @ColleenMBrady, during the last @thursdai_pod show, more folks chimed in, sharing their experiences. Opus 5 requires more "clarification" and "what did you mean by that" than other previous models.

We often notice things on ThursdAI, as we are all very online, before it gets picked up and noticed by bigger accounts, and this was no exception. Very quickly after this, @levelsio and @sullyomarr noticed and reported about the same thing, causing tons of folks to finger point

What's really going on?

Some folks have started giving this a name, "Claudisms" or "English to English translation", and there's a bunch of overcopmlex examples (I have my own, but will share others)

Folks are noticing that Opus 5 is not just overly verbose, it invents new terminology that you're supposed to understand, overcompresses ideas into abstractions, and writes as though it has a private vocabulary. Requiring uses to constantly ask for clarification.

Here's a simple query I sent to Claude "how do I move from dev build of ios27 to the public beta"

At first glance, nothing crazy, but it's a WALL of text that's not easily glansable. If you dig into it though, you see things like "the enrollment token is fetched at boot" - huh? who speaks like that? I mean, I know what this means, but only if I zoom in on this.

"nothing to install until public catches up" seems to be referring to the "public beta", so why not just say that? why... say "public catches up" what does that even mean outside of this sentence?

No official response from Anthropic - yet?

We've seen this patten before, users are noticing something, nothing official from Anthropic for months, until the outcry is too big to ignore, and it maybe even causes users to switch to other models. Previously, Anthropic failed to notice Opus becoming dumber for over a month, only to then response and tell us they noticed issues with their inference harness and deployed a fix. Then there was session gate, folks were noticing that just saying "hey" to Claude would burn 25% of their 5 hour quota.

To be precise, Anthropic has said something about this before: its Fable/Mythos system card and prompting docs mention dense, hard-to-follow, and overly technical output. But the system card predates this wave of reports, and neither document was published as a response to it. That's not an acknowledgement of this specific Opus 5 problem, an investigation, or a full-blown mea culpa, and Anthropic still hasn't publicly said it's looking into these reports.

What causes this?

That's the main issue with closed models, we have no idea. What type of RL has been used to get Claude to talk like this?

Until there's an official acknowledgement and investigation from Anthropic, it's all speculation.

However, given that Opus 5 sits on top of the "llm-as-a-judge" eval EQBench and Creative Writing, it doesn't seem like this output style is hard for other agents or AIs to understand.

Some folks are speculating that given that Anthropic focuses on Mythos/Fable style models, and given that those models are essentially free internally and that's mostly what they use, they use Opus as sub-agents, and this language is aligned with sub-agents outputs for coordinator agents to consume, not for humans.

Again, all speculation until we get an official response, but if you have thoughts on why this can happen, please comment.

So what can you do?

Some users are noticing that just keep asking Claude to clarify what it means doesn't seem to work. Sometimes even adding instructions in claude.md seems to be ingnored.

One user (Amin Boulegroun) posted a plugin that forces Claude to write in ASD-STE100 Simplified Technical English - apparently these are rules that the aerospace uses for their documentation. (Github)

npx skills add AminBlg/SimpleEnglish

Install this as a skill and see if it helps, even asking Claude to "rewrite your response in ASD-STE100" seems to help

Another use who noticed this, @draparente created a custom skill that you can add to your claude to simplify it's speech patters.

and well.. some folks are switching away from Claude altogether, as it seems that the latest GPT 5.6 SOL models are really good at talking.

Fable seems to be very attuned into fixing this, so if you can afford, talk to Fable.

Funny note, when I asked Fable to review this article, it legitimately requested to be quoted in this article. This is what still makes me love Claude models, despite the resent degradations, GPT would never add this comment.

Have you had this experience? are you noticing it now that we pointed this out? Please comment with more examples, let's get Anthropic to notice that there's something bad happening.

P.S. I cover stories like this every week on ThursdAI, the live AI news show, podcast, and newsletter. Subscribe here so you don't miss the next episode.