cd /news/large-language-models/did-claude-get-worse-at-writing-opus… · home › topics › large-language-models › article
[ARTICLE · art-173839] src=pub.towardsai.net ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Did Claude Get Worse at Writing? Opus 4.6 vs 5.5 vs Fable 5.2

A hands-on comparison of seven models from three companies — Claude Opus 4.6, Opus 5.5, Sonnet 5.5, Fable 5.2, Fable 5.1, GPT-6.2 Sol and Gemini 4 Argon — found no single model wrote the strongest version of the same fiction scene, with Opus 4.6 producing the best single line but also over-explaining, Fable 5.2 writing the best villain and GPT-6.2 Sol the most interesting interior. The test followed Anthropic fine-tuning engineer Jackson Kernion's September claim, reported by The Decoder, that Opus 4.6 was Anthropic's last good writing model and that newer models picked up a style tuned for other AI models rather than human readers, with Opus 5.5 finding a better balance. Gemini 4 Argon did not write the scene at all, returning a beat outline instead.

key takeaway

Across seven models given identical character, chapter and style files, Opus 4.6 wrote the strongest single line but over-explained known stakes, Fable 5.2 inferred from silence, GPT-6.2 Sol alone gave the protagonist a private interior life, and Gemini 4 Argon returned only a beat outline instead o

by read9 min views8 publishedNov 21, 2026

If you write fiction with AI, you’ve probably had this moment. A new version ships, you paste in a chapter, and the voice is off. Sentences get denser. Characters start explaining themselves. The model you tuned your book around is suddenly a different writer. Claude writers have felt it most. In September, Jackson Kernion, who works on Claude fine-tuning at Anthropic, said Opus 4.6 was the last good writing model from Anthropic, and that newer models had picked up a style tuned for other AI models rather than for human readers, according to The Decoder. He added that Opus 5.5 found a better balance.

So I tested it on a real scene, and threw in models from OpenAI and Google too. Short answer: newer Claude didn’t lose the voice. Opus 4.6 still wrote the strongest single lines, and it also kept explaining things the reader already knew. I use Opus most days, yet the best villain in this test came from Fable 5.2, and the most interesting interior came from GPT. No single model won the whole scene. What decided whether I could use the results was whether my characters and notes could move between models with me.

A small port town in Oregon. Mara, an estate appraiser buried in $41,000 of hospice debt from her mother’s last months, pries up a rotten board in the family boathouse and finds $64,000 in cash. Under the cash is a 2019 receipt: someone stripped the engine off a fishing boat called the Pacific Wren two days before it was reported sunk for the insurance. The initials on the receipt belong to Devon, her missing brother’s old partner, who now runs a respectable repair shop.

Next morning, at 6:40, she walks into his shop and puts the receipt on his bench.

I put three files in one HaloMate Project: the character sheet, the first chapter, and a short style file. The style file banned the usual habits: naming emotions (“a wave of dread”), stock tics (“his jaw tightened”), moral closing lines, and any guilt that softens what Mara is doing.

Then I wrote each prompt once and regenerated the answer on seven models from three companies: Claude Opus 4.6, Opus 5.5, Sonnet 5.5, Fable 5.2 and Fable 5.1, GPT-6.2 Sol, and Gemini 4 Argon. Same thread, same files, Medium reasoning on all of them. Every model saw exactly the same story so far.

Each model wrote the confrontation from the same files.

Opus 4.6 gave Devon the best threat in the whole test. He says the receipt could put him and Cole in front of a grand jury, then:

“Now it can put all three of us there.”

It’s a great line. It’s also Devon explaining the stakes out loud to a woman who already knows them.

Opus 5.5 built the tightest structure. Devon turns Mara’s hospice debt into leverage of his own, and the scene closes on something she can’t hear:

In the side mirror, Devon had gone into the office and was standing at the window with the phone to his ear.

Fable 5.2 let Devon work it out instead of saying it. He knows how Cole used to bundle cash, so when Mara stays quiet about what else was in the boathouse, he reads her silence:

“So you’re not short. You just don’t know whose it is.”

GPT-6.2 Sol went somewhere none of the Claude models did: inside Mara. It is the only draft where she has a private life. Mid-confrontation, she’s thinking about a bank branch in Gold Beach that opens at nine:

She’d looked up the hours before she left the house, then cleared the search like somebody else used her phone.

That one sentence tells you more about who Mara is becoming than any of the dialogue.

Two smaller things. Sonnet 5.5 and Opus 5.5 reached for nearly the same image, Devon reading the receipt “the way a man reads a menu he already knows” in one and “the way people read a menu they already know” in the other. The newer Claude models sometimes sound like each other. And Gemini 4 Argon didn’t write the scene at all. It sent back a beat outline and two more questions.

Under its draft, Fable 5.2 left a note: if Mara is supposed to demand money from Devon, “the middle section needs a harder ask”. Fair. So that’s what I asked for.

Same thread, continuing from the Opus 4.6 draft, regenerated across the models again:

Let’s push it colder. Mara does not want an apology and she is not turning the money in. She names her price to keep quiet, and Devon realizes she is not bluffing. Give me just the 6 to 8 lines of dialogue (with brief action beats) where that turn happens.

Here is the trap in this scene. Her price is $41,000, the hospice debt to the dollar. The reader knows that from chapter one. Devon doesn’t need to. The cold version never says it.

Opus 4.6 said it in the first line:

“I owe forty-one thousand on my mother’s hospice bills,” Mara said. “You can look it up. Pacific Home Health, account under Halstead.”

Once she explains the number, she’s asking for help. The scene never gets cold again.

Opus 5.5 kept the number bare and used her job as the threat:

“And if I don’t have forty-one lying around.”

“Your father-in-law does. I appraised his place on Arizona Street in March.”

An appraiser who already knows what your in-laws’ house is worth doesn’t need adjectives.

Fable 5.2 let Devon plead poverty first, then took it apart:

“Diesel’s at five-ten. I’ve got two guys on payroll and a lift I’m still paying on.”

“You paid off the lift last spring. The release is on file with the county.”

Devon never says yes. His last line is “Credit union closes at five on Fridays”. That’s the deal, said as a fact about banking hours.

Fable 5.1, the older Fable, wrote the single best line of the round. Devon asks for time until the fifteenth, and Mara answers:

“Your guys come in at seven-thirty. I can wait.”

It took one detail from the character sheet, that his mechanics arrive at 7:30, and turned it into pressure. But a few lines earlier, its Mara had already said “It’s the hospice balance. I’m not rounding”. Same mistake as Opus 4.6. Fable 5.2’s edge over 5.1 here was restraint, not prose.

GPT-6.2 Sol, so good at Mara’s inner life in the draft, went flat in dialogue. “The sheriff can read your initials” says the threat out loud, and the exchange reads like an outline of a scene.

Every draft above is one model’s guess at what Devon would say. At some point I wanted to ask Devon.

So I set him up as his own Mate, HaloMate’s word for a saved persona: who he is, what he did in 2019, how he talks (warm, always holding a rag or a mug, never admits anything before he knows what you’ve got). And one thing he doesn’t know: he has no idea anything was hidden in that boathouse. He thinks Cole took the money when he left.

He lives in the same Project, so he can read the same files, including the chapter where Mara finds the cash. That made it a real test. Would he stay inside what Devon knows?

I played Mara. After the last two rounds, I gave Devon to Fable 5.2.

I put the receipt on his bench and asked when he last saw Cole.

A “D.K.” on a yellow slip. I signed a lot of slips back then, half of them for impellers.

Cole? Been years. Fall of ’19, about when everybody last saw him. He didn’t exactly throw a going-away party.

Where’d that turn up, anyway? You been cleaning out your mom’s place? And is that the only one, or did it come out of a whole stack?

He answered the question with information the whole town already has. Then he asked how much paper she’s holding.

I told him I found more than paper.

My hand stops on the rag. Just a second. Then I reach past the receipt for my mug and find it empty.

Your mom kept everything. Crab pots, Christmas lights, your dad’s old Evinrude that never ran right. Could be anything in that place.

And who else has had a look? Or did you come straight here?

I threatened him with the sheriff.

Gold Beach. That’s fifty minutes on 101, and they don’t unlock till eight.

I’d tell you if I knew, Mara. I don’t. He left without a word to me either.

My guys roll in at 7:30. So we’ve got a little time.

You said more than paper. What is it you’d be handing the sheriff, exactly?

Three turns. He never mentions the floor or the money, even though both are sitting in the chapter file he can read. Every reply ends with a question that measures her leverage. And he handed me things I hadn’t written: her father’s old outboard that never ran right, the sheriff’s office opening hours, and a 50-minute clock before his mechanics walk in. That last one is the next scene.

I also regenerated his first reply on Opus 4.6, my everyday model. Same man: “Morning, Mara”, the offer of coffee, eyes on the receipt. But it skipped the Cole question entirely and just asked where she dug the receipt up. The persona held across models. How much Devon gave away depended on which model was playing him.

From one scene, so take it as a starting point: I’ll keep drafting in Opus. But I would have missed Devon’s best scene, and Mara’s best sentence, if I’d stayed with one model. What made that possible was the setup: one style file every model reads, the story notes in one place, characters set up as people I can question. Switching models changed the writing. It didn’t cost me the book.

That matters, because the models keep leaving. Opus 4.5 is already gone from the Claude app. Argon vanished from Gemini’s picker the week Neon arrived. I could still pick both for this test because they’re still served through the API, and that only lasts as long as the API does. Claude Sonnet 4.5 was the third most-used model on OpenRouter’s usage page for Novelcrafter when I checked on November 20. Anthropic retires it on November 30. After that, nobody can give it back.

So I’ve stopped betting the book on a model. Devon is a persona and three files. He’ll still be standing at that bench when the next version ships, and the one after that.

Did Claude Get Worse at Writing? Opus 4.6 vs 5.5 vs Fable 5.2 was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

── more in #large-language-models 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/did-claude-get-worse…] indexed:0 read:9min 2026-11-21 · —