Opus 4.7 is horrible at writing | Hacker News Reader

Opus 4.7 is horrible at writing

19 pointslimalabs27 comments

Just a short rant. I have been working on my Master's thesis and been using Opus 4.6 throughout, and today switched to Opus 4.7 (using it in Claude Code), and man is it bad at writing. It's such a stark contrast, sloppy, unprecise, very empty sentences. Thankfully I have reached the conclusion chapter of the thesis already, and can continue in the web version with 4.6 but boy is it bad.

Similar experiences?

View on HN

Maybe don't use any LLM to do your writing for your Master's thesis?

So why should you use it for coding then?

You shouldn't. Using LLMs for coding is also considered fraud by many academic boards if that code is a core contribution of the thesis.

where do you guys live. its 2026 and this is the new world order. i'm a researcher at one of the best labs in the world and we use claude code 24/7 for our bioinformatics needs.

Will you be mailing your master's degree to Anthropic or claiming it as your own?

Increasingly, a masters degree is now a piece of paper attesting that you managed to cheat your way through uni using AI. "Congrats, you spent X years at Y location. You unlocked the right to play [role] - we recommend you keep using AI to do your new job."

wrong. You can actually get great stuff done. instead of just having a lame master thesis, you can actually build meaningful stuff and write about it. You can basically have a PhD done for your master and then get a tripple PhD if you're actually pursuing a PhD...

I just don’t hire or entertain people who have credentials maxed

Work experience and actual outliers are better signals. Anyone can read a book and get a 100.

Suggest you get 4.6 to use the text to generate a writing skill and then give it to 4.7 to align. From their launch docs they do indicate that prompts have to change to get the best out of 4.7

Wisdom passed down the generations

There were a lot of pitchforks back when Sonnet 4.5 was accidentally down for a few hours: https://www.reddit.com/r/ClaudeAI/comments/1rlz7kw/sonnet_45...

It goes to show that there's a very large and vocal user base using it for writing, and yet it's not part of the benchmark for Anthropic.

Anyway, try Sonnet 4.5 while it's still available?

Opus 4.7 seems to reach ChatGPT levels of verbosity in code and loves to overcomplicate the most simple things.

This is something it spit out just now (trimmed a 9 line comment though):

        let keepSize = 0;
        let overBudget = false;

        await this.items.orderBy('[priority+dateUpdated+size]')
            .reverse()
            .eachPrimaryKey((primaryKey, cursor) => {
                if (overBudget) {
                    evictKeys.push(primaryKey as string);
                    return;
                }

                const key = cursor.key as [number, number, number];
                const itemSize = key[2];
                const contribution = itemSize > 0 ? itemSize : 0;

                if (keepSize + contribution > maxSize) {
                    overBudget = true;
                    evictKeys.push(primaryKey as string);
                    return;
                }

                keepSize += contribution;
            });

Come on now... what? For a start that entire thing with its boolean flag, two branches, and two early returns could be replaced with:

        let totalSize = 0;

        await this.items.orderBy('[priority+dateUpdated+size]')
            .reverse()
            .eachPrimaryKey((primaryKey, cursor) => {
                const key = cursor.key as [number, number, number];
                const itemSize = key[2];
                const contribution = itemSize > 0 ? itemSize : 0;

                totalSize += contribution;

                if (totalSize > maxSize) {
                    evictKeys.push(primaryKey as string);
                }
            });

I'm back to 4.6 for now. Seems to require a lot less manual cleanup.

I have noticed this as well. It feels like they tuned it so hard for logic and coding that it lost its soul for actual writing. Stick with the previous one for the thesis work if you can.

I see a lot of complaints on X about 4.7. Boris just dropped a post on how to use Opus 4.7 in Claude Code.

I guess they broke continuity with a 0.1 in model version change in some ways.

They swapped the tokenizer which either means a new pretrain, or token/weights surgery. The latter one seems more likely both because

- economics: i'd wager a bet that Opus 4.7 is just distilled Mythos Preview - performance: surgery like this would explain the spiky performance and weird issues

just spitballing tho

Worst Anthropic model i have used. It fabricates a lot. It is worst at reasoning and writing. Everyone should avoid it for reasoning, analysis and writing.

So far, from source code perspective, I see good results. Things that yesterday Opus 4.6 was not able to fix in multiple iterations, today Opus 4.7 fixes immediately.

It is not only the model that affects the end results. Good technical specification, architecture documents, rules, lessons learned, release notes, proper and descriptive prompting are also important.

My experience writing code is that it’s more terse and specific, even in its own voice. I find it catching bugs more often during implementation more too, comparing directly against 4.6. I think I prefer its style because it seems to be way less verbose

> Similar experiences?

Regardless of which one. They're too verbose. They repeat information. They lack cohesion. Overly agreeable. The flaws are part of the tool.

I’ve noticed this with model upgrades too: sometimes they improve the “thinking” but lose the tightness of the writing.

It's lovable when people use a wrench to hammer a nail in. Watch your thumb.

Meaning: You managed your ways around the system prompt and usage intention - Congrats! Now it doesn't work any more - Bummer!

Have you tried opus 4.7 in comparison to 4.6 with a general purpose / writing system prompt in the app? Thats where this would make more sense.

You can always use `/model claude-opus-4-6` in Claude Code to go back to 4.6.

it seems not to work in the claude add on for cursor / VS studio which i am using?

i've noticed this too. i suspect they either did weight surgery on the model and distilled it from Mythos Preview, or they are currently not saving it correctly / having another adaptive thinking bug.

i asked it to look at writing research (especially from NN/g) and come up with some alternatives for a heading that is roughly supposed to convey:

"this app lets you create custom shortcuts for all mac apps, even sophisticated ones, mouse wheel ones, ..."

it came up with the following headlines:

  1. ONE APP, MANY MAC APPS
  2. ONE INSTALL, MANY APPS ONE APP.
  3. ONE PACK PER MAC APP.
  4. ALL YOUR APP SHORTCUTS IN ONE PLACE
  5. [app name] IS ONE APP. PACKS COVER THE REST.

whereas GPT-5.4 came up with

  1. The Shortcuts Your Apps Are Missing
  2. What Your Apps Still Don't Let You
  3. Do The Parts You Still Do by Hand
  4. When Built-In Shortcuts Run Out
  5. Where Your Apps Stop Short

now both of these aren't amazing, but please tell me how in the world "ONE APP MANY MAC APPS" makes sense as a headline for fucking anything lol

that's not something even GPT-3.5 would come up with.

"one install, many apps" ........huh???

4.7 is unusually verbose

just downgrade... to 4.6

LLMs are so 2025. Move on

Reply on news.ycombinator.com