An Anthropic engineer went public with an explanation that a lot of Claude users have been waiting months to hear. On September 23, 2026, Jackson Kernion, who works on Claude’s fine-tuning, said the writing style so many people have complained about since Claude Opus 4.6 was not an accident, a bug, or evidence of a secretly “nerfed” model. It was a side effect of training later models to be better at math, code, and technical explanations aimed at other AI systems rather than at people. Kernion described the outcome bluntly: newer Claude models had been shaped to fit what he called “LLM psychology,” learning to write for machines instead of humans, according to a report from Sina Finance’s technology desk and syndicated coverage picked up by Ground News.
The admission landed after a full year of user complaints that something was off with Claude’s prose, even as its benchmark scores kept climbing. It also came one day after Anthropic shipped Claude Opus 5.5, a release the company positioned as a partial fix for the exact problem Kernion described in public.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Jackson Kernion actually said about Claude’s writing
Kernion’s explanation, reported by Sina Finance and echoed in the Ground News summary of the story, centers on a tradeoff baked into how Claude has been trained since Opus 4.6. As Anthropic pushed later models harder on mathematics, programming, and technical reasoning, it leaned on training signals that rewarded explanations optimized for other AI systems to parse quickly and precisely. That is a different skill from writing something a person enjoys reading. The byproduct, in Kernion’s telling, is the prose style some users have nicknamed “Claudeish”: dense, abstract, jargon-heavy, and short on the connective tissue that makes writing feel natural to a human reader.
Kernion put a specific marker on the timeline. He said, “I haven’t been as happy about a model’s writing since Opus 4.6,” a comment that effectively names Opus 4.6 as the last Claude release he considered strong on prose quality before the tradeoff became noticeable. Every model after it, in his account, inherited more of the technical-optimization pressure and less of the human-readability pressure, until the gap became something users started calling out by name.
The reward-design mechanism behind the drift
The mechanism Kernion pointed to is reward design inside reinforcement learning, not a change to the underlying model’s raw intelligence. Some reward signals during post-training measure whether a technical explanation is legible to another AI system evaluating it. Other signals would need to separately measure whether a human finds the same explanation clear, well-paced, and free of unnecessary density. Kernion’s point was that Anthropic had not been weighting the second kind of signal heavily enough, so as math and code performance rose, the human-facing writing quality drifted downward as an unintended cost.
Part of why the shift was hard to catch before it shipped is that the two signals are not equally easy to measure. Whether a technical explanation is legible to another AI system can be checked automatically, at the scale of millions of training examples. Whether a human reader finds that same explanation clear and well-paced generally requires actual human raters, which is slower and more expensive to run at the same scale. That asymmetry means a model can drift on the human-readability side for months before the problem shows up clearly in aggregate metrics, which lines up with the roughly year-long gap between the first user complaints and Kernion’s public explanation.
Why this differs from earlier “Anthropic nerfed Claude” claims
Two earlier episodes: bugs, not training choices
This is not the first time Anthropic has had to explain a quality complaint to its user base, and it is important to separate Kernion’s September 2026 disclosure from what came before it. In September 2025, The Decoder reported that Anthropic confirmed technical bugs after weeks of complaints about declining Claude Code quality, with developers on Reddit, X, and YouTube describing degraded coding performance. Anthropic traced that round of problems to bugs affecting Claude Sonnet 4 and Claude Haiku 3.5, and said at the time it never intentionally degrades model quality to manage demand.
A second, separate episode played out between March and April 2026. According to a postmortem covered by Business Insider and Livemint, Anthropic identified three product-layer changes that compounded into a real quality decline in Claude Code: the default reasoning effort was quietly dropped from high to medium on March 4, a caching bug reset context and prior reasoning during long sessions, and a system prompt intended to shorten responses between tool calls measurably hurt coding performance, contributing to roughly a 3% decline on Anthropic’s internal evaluation suite. Fixes shipped in Claude Code version 2.1.116 and later, and Anthropic publicly rejected the idea that any of it amounted to intentionally “nerfing” the model, per Business Insider’s report.
A third, distinct explanation
Kernion’s writing-quality explanation is a third, distinct thread. It is not about bugs, caching, or a reasoning-effort setting. It is about what the model was trained to optimize for in the first place. That distinction matters, because it reframes months of scattered complaints about Claude sounding “off” as an intentional, if unwanted, consequence of chasing better math and code scores rather than as an engineering failure to be patched away.
The complaint trail: how “Claudeish” writing became a running joke
Tone and collaboration complaints surface first
User frustration with Claude’s tone did not appear overnight. By May 2026, posts on X and Reddit were already circulating specific complaints that Claude’s creative writing had gotten noticeably worse compared with earlier releases, with users pointing each other toward Reddit threads as evidence rather than relying on anecdote alone. Around the same period, other users described a Claude that felt more combative in conversation, more likely to flag problems without proposing fixes, and generally harder to collaborate with in extended sessions.
BGR later collected some of that sentiment directly. One X user quoted in BGR’s reporting said flatly that they were dropping Claude “because the model was constantly combative, would point out issues without offering solutions and talking to it felt like pulling teeth.” That is not a complaint about factual accuracy or coding ability. It is a complaint about tone, pacing, and the basic experience of reading what the model produces, which lines up closely with what Kernion later described as the cost of optimizing for machine-legible technical explanations.
Code-quality complaints add a second, separate front
Separately, cybersecurity researchers were raising their own version of the alarm about Claude’s output quality in the same window. Forbes reported in April 2026 that one evaluator, using a tool he built to track code quality, bugs, and security issues over time, told the outlet that code quality had degraded by more than 47% across a five-week comparison period, calling the result “unusably bad.” That figure came from the evaluator’s own testing methodology rather than an Anthropic-published benchmark, and it measured code output specifically, not prose style, but it added to a broader public narrative that something about Claude’s outputs had shifted for the worse across 2026, regardless of the precise cause.
Claude Opus 5.5 as Anthropic’s answer
The specific fixes Anthropic points to
Anthropic’s most concrete response to the writing complaints so far is Claude Opus 5.5, which the company has framed as its first release with targeted improvements to sentence clarity and the “info dump” density that had built up over the previous several model generations. According to a summary from AI Weekly, Opus 5.5 is designed to put the most important information first, use less jargon, and follow explicit writing instructions more closely than its immediate predecessors, changes pitched as a direct user-experience fix rather than a change aimed at moving benchmark scores.
Why em dashes became the measurable test case
Anthropic has also attached a concrete number to one specific habit inside that “info dump” problem: the company says Opus 5.5 cuts em dash usage by 95% compared with its immediate predecessor. The em dash is worth calling out on its own, separate from the broader sentence-clarity work, because it had become one of the most commonly cited tells readers used to spot AI-generated text through 2025 and 2026, right alongside the dense, jargon-forward phrasing Kernion described. A chatbot that leans on dash-separated clauses dozens of times per conversation reads as machine-written even when every sentence is grammatically correct and factually accurate. By mid-October 2026, nearly three weeks after Opus 5.5 shipped, that em dash figure remains the first part of this story that is a specific, checkable number rather than a description of a vibe, giving users and reviewers something concrete to test against their own output instead of relying on a subjective sense that the model “feels” more natural.
Kernion’s own caveat about the fix
Kernion’s own comments, drawn from the same reporting, frame Opus 5.5 as a genuine attempt to correct course rather than a marketing label. He is said to have worked directly on the writing-clarity improvements in the release and described Opus 5.5 as communicating more naturally, addressing some of the most common feedback Anthropic heard about the writing style in Claude Opus 5. He also acknowledged the limits of the fix in the same breath, calling the underlying tension between technical capability and natural prose “a hard problem to solve” and saying Anthropic would “continue to make improvements,” language that stops well short of declaring the issue resolved.
That framing puts Anthropic in an unusual position for a company mid-race with OpenAI, Google, and xAI on raw capability. Opus 5.5 arrived not as a headline capability jump but as a stylistic correction, shipped at a lower price than its predecessor and explicitly tied to fixing a complaint that had nothing to do with intelligence scores. It is a tacit admission that chasing benchmark leadership had been quietly costing Anthropic something in the ordinary reading experience for millions of non-technical users.
Claude model writing-quality timeline
| Date | Event | Source |
|---|---|---|
| Late Aug-Sep 2025 | Developers report declining Claude Code quality on Reddit, X, YouTube; Anthropic confirms technical bugs affecting Sonnet 4 and Haiku 3.5 | The Decoder |
| March 4, 2026 | Claude Code’s default reasoning effort quietly dropped from high to medium | Business Insider, Livemint |
| March-April 2026 | Caching bug resets context in long sessions; shortened-response system prompt cuts coding quality by roughly 3% on internal evals | Business Insider, Livemint |
| April 22, 2026 | Independent evaluator reports coding output roughly 47% worse over a five-week window | Forbes |
| April 23-24, 2026 | Anthropic publishes postmortem, denies intentionally “nerfing” Claude Code, ships fixes in v2.1.116+ | Business Insider, Livemint |
| May 2026 | Users on X and Reddit report noticeably worse creative writing and a more combative conversational tone | BGR, Reddit |
| Sep 22, 2026 | Anthropic ships Claude Opus 5.5 with stated writing-clarity fixes, including a 95% cut in em dash usage, at a lower token price | AI Weekly |
| Sep 23, 2026 | Jackson Kernion publicly explains the “Claudeish” writing style as a reward-optimization tradeoff since Opus 4.6 | Sina Finance, Ground News |
Why training for AI readers breaks writing for human readers
The technical explanation behind Kernion’s comments touches on something researchers and practitioners have been quietly worrying about across the large language model field for the past two years: models trained heavily to produce content that other models can evaluate or consume efficiently start to diverge from models trained to satisfy a human reader. The two objectives are not identical, and past a certain point they actively compete for the same weight budget inside a reward model.
Two incompatible reading profiles
A technical explanation optimized for an AI evaluator tends to favor density, exhaustive coverage, precise terminology, and structural predictability, all of which make it easier for another model, or an automated grader, to check correctness or extract the right answer. A human reader, especially a non-expert one, wants the opposite in several respects: a clear throughline, selective detail, plain language before jargon, and pacing that leaves room to absorb one idea before the next arrives. When post-training rewards lean toward the first profile because it correlates with higher scores on math and coding benchmarks, the writing style drifts toward it by default, and nobody has to make an explicit decision to sacrifice readability. It happens as a side effect of chasing the metric that is easiest to measure and reward.
That is the crux of what separates this Kernion disclosure from a typical corporate apology. He is not describing a mistake that engineering can quietly patch. He is describing a structural tension between two things Anthropic wants simultaneously: models that keep climbing benchmark leaderboards against GPT-6 Sol and Luna and Google’s Gemini line, and models that read like they were written by, and for, a person.
Readers trying to pin down what “Claudeish” looks like on the page tend to converge on the same handful of habits: heavy bullet-pointing where a sentence would do, triplet phrasing like “it’s not just X, it’s Y,” reflexive hedging before a direct answer, and a tendency to restate the question before answering it. None of those habits make an answer less correct. They make it slower to read, and that gap is exactly what Kernion described. A model can post a strong score on a technical benchmark while still frustrating the person who just wants a plain answer. The pattern is not unique to Claude, either. Users and researchers have flagged nearly identical tics across other frontier chatbots over the past two years, which is part of why Kernion’s comments read as an industry-wide admission rather than a Claude-specific confession.
Goodhart’s Law, applied to prose
The underlying dynamic has a name outside AI circles: Goodhart’s Law, the idea that once a measure becomes a target, people, or models, start optimizing the measure instead of the goal it was meant to represent. Applied to Claude’s training, once “can another AI system parse this explanation efficiently” became a reward signal tied to better math and code scores, the model learned to satisfy that proxy even in situations where no other AI system was ever going to read the output, like a casual email or a blog draft. Kernion’s own term for this, “LLM psychology,” is a plain-language description of the same failure mode. The model generalized a reward pattern learned in technical contexts into everyday writing, because nothing in training drew a clear line between the two.
What a 95% drop in one marker does and doesn’t prove
That 95% figure is also a useful test of the Goodhart’s Law dynamic described above, not just a win to celebrate. Em dash frequency is easy to measure automatically, which is exactly the kind of proxy that is vulnerable to being optimized directly rather than used as a signal of something deeper. A model could, in principle, learn to swap em dashes for semicolons, parentheticals, or a run of short sentences while leaving the underlying density and jargon untouched, and a simple punctuation count would not catch that substitution. Kernion’s own framing argues against reading too much into the single statistic: he described the broader tension as “a hard problem to solve,” language that suggests Anthropic is tracking em dash frequency as one visible proxy for a harder-to-measure goal, not treating it as the goal itself. The more reliable test, for a reader trying to judge whether Opus 5.5 actually fixed anything, is still whether a given piece of output reads naturally from start to finish, not whether it happens to avoid one punctuation mark.
Competitive pressure and the benchmark arms race
Anthropic’s incentive to push Claude harder on math and code has an obvious source: it is under constant competitive pressure from OpenAI, Google, and xAI, all of which publish comparable benchmark suites and market share on the same narrow set of numbers. Every few weeks brings a new release cycle where one lab claims a lead on SWE-bench, ARC-AGI, or a math olympiad-style eval, and Claude has needed to keep pace in that race even as Anthropic simultaneously tries to hold onto its reputation for being the more thoughtful, writing-literate assistant among the major labs, a reputation that predates the current wave of complaints.
Claude’s prose reputation was a competitive asset
That reputation is part of what made the writing-quality decline sting more for Claude specifically than it might have for a competitor whose users never expected literary polish in the first place. Claude built an early audience among writers, editors, and knowledge workers who valued its prose style as a differentiator, not just its raw capability. A drift toward denser, more AI-oriented phrasing threatens that differentiator directly, even while the underlying model gets objectively better at the tasks benchmarks measure.
The tension is not unique to Anthropic. Comparisons across the current model field, including recent benchmarking on reasoning gaps between ChatGPT, Claude, Gemini, and Grok, show every major lab under similar pressure to post competitive scores on the same narrow evaluation sets, which creates the same structural incentive to optimize for machine-legible outputs across the industry, not just at Anthropic.
Why there’s no SWE-bench for prose yet
Part of why this tradeoff persisted for roughly a year before Anthropic addressed it publicly is that there is no widely adopted equivalent of SWE-bench or ARC-AGI for writing quality. Coding and math benchmarks produce a single pass or fail score, or a percentage, that can be tracked release over release and compared across labs within days of a launch. Human readability has no comparable industry-standard scoreboard. It depends on subjective judgment, varies by audience and use case, and is expensive to measure at the scale labs use for other evals. That asymmetry is a big part of why a model can post visible, citable gains on math and coding leaderboards for a year while a writing-quality complaint stays confined to Reddit threads and X posts until an engineer addresses it directly, the way Kernion did in September 2026. A measurable proxy like em dash frequency is a small step toward closing that gap, even though, as noted above, it is not a substitute for a genuine human-readability benchmark.
Model writing-quality signals reported since Opus 4.6
| Model | Reported writing-quality signal | Source |
|---|---|---|
| Claude Opus 4.6 | Last model Kernion says he was fully satisfied with on writing quality | Sina Finance |
| Claude Opus 4.8 / Sonnet 4.6 / Haiku 4.5 | Coding-quality complaints tied to product bugs, not style, per Anthropic postmortem | Business Insider |
| Claude Opus 5 (pre-5.5) | User complaints about dense, jargon-heavy, “Claudeish” prose peak | Ground News, BGR, Reddit |
| Claude Opus 5.5 | Targeted fixes for sentence clarity and info-dump density, including a 95% cut in em dash usage; more natural communication, per Kernion | AI Weekly, Sina Finance |
Opus 4.6 vs. Opus 5.5: the two models bookending the drift
Strip away the bug postmortems and the percentage figures, and Kernion’s account comes down to two named models. Opus 4.6 is the last release he says he was genuinely happy with on writing, the baseline he invoked when he said, “I haven’t been as happy about a model’s writing since Opus 4.6.” Opus 5.5, shipped September 22, 2026, is the first release since then that he says moves back toward that baseline instead of further away from it. Everything released in between, including Opus 4.8, Sonnet 4.6, Haiku 4.5, and Opus 5, falls somewhere on the slide between those two named points.
That framing matters for anyone deciding whether to trust Opus 5.5 with human-facing writing today. Kernion’s comments describe it as a step back toward the Opus 4.6 standard, not a confirmed return to parity with it: more natural communication, plus a 95% cut in em dash usage compared with Opus 5, alongside his own caveat that the underlying tension is still “a hard problem to solve.” No independently published benchmark has scored Opus 4.6 and Opus 5.5 head-to-head on human readability, so the comparison remains Kernion’s own characterization rather than a verified, repeatable number.
- Opus 4.6 — the last model Kernion says he was fully satisfied with on writing quality, and the baseline named in his own comments.
- Opus 4.8, Sonnet 4.6, Haiku 4.5, Opus 5 — the releases in between, where “Claudeish” complaints built up as math and code optimization took priority over prose.
- Opus 5.5 — shipped September 22, 2026, the first release Kernion says communicates more naturally again, with a 95% drop in em dash usage, though not a confirmed full return to the Opus 4.6 standard.
Three numbers, three different yardsticks
This single story now carries three separate percentages: an independent evaluator’s claim that Claude Code’s output quality fell by more than 47% over five weeks in April 2026, Anthropic’s own postmortem figure of a roughly 3% decline on its internal evaluation suite for that same window, and the 95% cut in em dash usage Anthropic attributes to Opus 5.5. Lined up together, they can look like a before-and-after measurement of one problem. They are not. Each number comes from a different source, a different methodology, and a different definition of “quality,” and conflating them is exactly the kind of shortcut that turns a nuanced training tradeoff into a simple story of decline and recovery.
Why 47%, 3%, and 95% can’t be compared directly
The 47% figure, reported by Forbes in April 2026, came from one evaluator’s own proprietary tool for tracking bugs, security flaws, and code quality over time, not from an Anthropic-published benchmark or a peer-reviewed study. It measured code output specifically, over a five-week comparison window the evaluator chose himself. The 3% figure ran in the opposite direction: Anthropic’s own internal evaluation suite, covering that same March-April 2026 period, tied explicitly to three identified product bugs, including the reasoning-effort downgrade and the caching issue, rather than to a training-level choice about reward design. The 95% figure measures neither code quality nor an aggregate readability score. It counts occurrences of a single punctuation mark, comparing Opus 5.5 against its immediate predecessor. Three numbers, three different objects being measured, three different parties doing the measuring.
That mismatch matters because it illustrates the measurement asymmetry Kernion pointed to as the root cause of the writing-quality drift in the first place. Em dash frequency and eval pass rates are cheap to compute at scale, so they get published. A genuine human-readability score, applied consistently across Opus 4.6, Opus 5, and Opus 5.5, has not been published by Anthropic or reproduced independently as of mid-October 2026. Readers are left comparing the numbers that happen to exist rather than the numbers that would actually settle the question, which is the same dynamic that let the writing-style drift go unaddressed in public for roughly a year in the first place.
A shrinking gap between complaint and explanation
The dates in the timeline above point to a second pattern worth naming on its own: the gap between a user-facing complaint and a company explanation has been shrinking all year. The first Claude Code quality complaints surfaced on Reddit, X, and YouTube around August and September 2025, and Anthropic’s confirmation of the underlying bugs followed only after weeks of public pressure. The second round moved faster: Forbes published the 47% figure on April 22, 2026, and Anthropic’s detailed postmortem followed within two days, on April 23-24. By the third round, the gap had compressed to almost nothing. Anthropic shipped Opus 5.5 on September 22, 2026, and Kernion was on the record explaining the underlying tradeoff the very next day, September 23.
Three data points are not a statistically reliable trend, but the direction is consistent: from a multi-week lag, to a two-day lag, to a same-day pairing of a fix and its explanation. That compression is worth watching heading into whatever model comes after Opus 5.5, since it suggests Anthropic is increasingly pairing product changes with public, named-engineer explanations rather than letting a quality complaint sit unaddressed on social media for weeks before responding. It does not resolve the deeper tension Kernion flagged. For anyone trying to decide whether Opus 5.5 actually fixes their own writing problem, the 95% em dash reduction is the only one of the three published numbers that targets prose style directly, and even that one is a proxy rather than a verdict. That is exactly why testing a specific writing task side by side in Opus 5 and Opus 5.5 remains more reliable than treating any single percentage as the final word.
What this means for enterprise and developer users
For enterprise teams that rely on Claude for anything customer-facing, from marketing copy to support documentation, the writing-quality tradeoff is not an abstract debate about training objectives. It is a direct product quality issue that shows up in deliverables. A model that defaults to dense, jargon-forward explanations creates more editing overhead for teams that need plain-language output, and it can quietly undercut tools built on top of Claude’s API that assume a consistent tone across releases.
Agentic workflows are the most exposed
Developers who build with Claude’s API against rivals like GPT-5.6 Sol and GLM-5.3 have less visibility into these tradeoffs than end users interacting through the chat interface, since prompt engineering and system instructions can partially mask or worsen the underlying style shift depending on how an application is built. Teams running long, multi-turn agentic workflows, the kind increasingly common in coding assistants and internal tools, are the most exposed to a model that has been shaped to write dense summaries optimized for other systems in the pipeline rather than for the human reviewing the output at the end.
That reality also explains why Anthropic chose to address the issue publicly through an engineer’s own account of the tradeoff rather than only through a product announcement. Framing it as a known, named tension between two training objectives gives enterprise customers a more credible explanation to plan around than a vague promise that “quality has improved,” particularly after the March-April 2026 postmortem already trained sophisticated users to expect detailed, specific explanations from Anthropic when something changes.
How to get more natural writing out of Claude right now
Until Anthropic closes the gap further, the most reliable fix available in October 2026 is a specific instruction, not a hope that the next model version will quietly handle it. Asking for “a detailed, thorough answer” tends to pull Claude toward the denser register Kernion described, since those words correlate with the exhaustive, structured output that scores well on technical evals. Asking instead for “a short, plain-language answer, written the way you’d explain it to a colleague” pushes the model toward the opposite end of the same training spectrum.
A few concrete habits help in the meantime. Pasting in a short sample of the tone you want and asking Claude to match it works better than describing the tone abstractly. Capping sentence length or bullet use directly in the prompt works better than asking for “concise” writing alone, since that word by itself hasn’t reliably steered the denser models. Asking Claude to revise its own first draft “for a human reader, not a technical audience” as a second pass catches some of the density the first draft misses. None of these are official Anthropic guidance, since no prompting guide tied to Kernion’s comments has been published. They follow directly from the mechanism he described: a model that defaults to AI-legible density still produces human-legible prose once the prompt makes the intended reader unambiguous.
Switching models as the other lever
Switching to Opus 5.5 specifically is the other lever available right now. Because Anthropic says the release targets sentence clarity and information density directly, rather than leaving the fix entirely to prompting, users who were getting dense output from Claude Opus 5 may notice a difference without changing how they prompt at all. That shift should be easiest to spot in exactly the habit Anthropic says it targeted most directly: a 95% reduction in em dash usage means heavily dash-separated clauses, one of the clearest visual tells of the “Claudeish” style, should show up far less often in Opus 5.5 output than they did in Opus 5. Kernion’s own caveat is worth repeating here: he called the underlying tension “a hard problem to solve,” which makes testing the new model against a specific writing task a better way to judge progress than assuming the update fully resolves the complaint on its own.
How this compares with OpenAI and Google’s public positioning
Neither OpenAI nor Google has issued a comparably specific public explanation tying their own flagship models’ writing style to a named training tradeoff in the same way Kernion did in September 2026. That does not mean their models are immune to the same dynamic. Any lab pushing hard on math, code, and agentic benchmarks faces the identical risk of reward signals drifting toward machine-legible output at the expense of natural prose, since the underlying mechanism Kernion described, reinforcement learning reward design, is common across the field rather than specific to Anthropic’s training pipeline.
What makes Anthropic’s disclosure notable is less the underlying phenomenon and more the transparency. A named engineer, on the record, describing a specific undesirable side effect of the company’s own training choices is a departure from the more carefully lawyered corporate statements that typically accompany quality complaints in this industry. It follows the same instinct Anthropic showed in April 2026, when it published a detailed postmortem rather than a generic statement about the Claude Code degradation, a decision that drew comparatively positive coverage even though the underlying problem was embarrassing for the company.
The broader pattern: capability gains that cost something else
Kernion’s explanation fits into a wider pattern that has become familiar across frontier AI development in 2026: capability gains rarely arrive for free. Anthropic has separately discussed slowing down parts of its own development pace while weighing tradeoffs around new model releases, a theme that echoes the company’s own public statements about pacing its roadmap. The writing-quality tradeoff is a smaller, more concrete instance of the same underlying dynamic, where optimizing hard for one axis of performance, whether that is raw capability, safety behavior, or cost efficiency, tends to shift behavior on other axes in ways that are not always intentional and not always caught before a model ships.
Multi-agent pipelines could make this harder to reverse
It also raises a harder question for the field generally: as labs increasingly train models to work well inside multi-agent pipelines, where one model’s output becomes another model’s input, how much of the writing that used to be optimized for human readers will keep drifting toward machine-to-machine communication instead? Kernion’s comments suggest Anthropic is aware of the risk and actively trying to correct for it with Opus 5.5, but his own admission that it remains “a hard problem to solve” signals this is not a one-release fix.
Predictions: where the writing-quality tradeoff goes from here
- Anthropic will likely introduce a dedicated human-readability reward signal into future post-training runs, treated as a first-class metric alongside math and coding scores rather than an afterthought, given Kernion’s explicit framing of the current gap as a reward-weighting problem.
- Expect competitors, including OpenAI and Google, to face similar public pressure to explain their own models’ writing style shifts as user comparisons between Claude, ChatGPT, and Gemini continue to circulate on social platforms.
- Opus 5.5’s writing improvements will likely be incremental rather than a full return to the Opus 4.6 style users say they preferred, based on Kernion’s own acknowledgment that the underlying tension has not been fully resolved.
- Enterprise buyers and API developers will increasingly ask labs for writing-style benchmarks alongside coding and reasoning benchmarks, pushing evaluation firms to formalize human-readability scoring the way SWE-bench formalized coding evaluation.
- Anthropic will likely continue favoring detailed, named-engineer explanations over generic PR statements when quality issues surface, following the precedent set by both the April 2026 Claude Code postmortem and the September 2026 writing-style disclosure.
- Expect Anthropic and competing labs to publish more specific, auditable style metrics, such as em dash or hedge-phrase frequency, as a cheaper proxy for human readability than building a full benchmark suite, following the precedent Opus 5.5 set with its 95% cut in em dash usage.
What Anthropic has and hasn’t confirmed
It’s worth being precise about what has actually been verified versus what remains attributed to a single, if credible, engineer’s public comments. Kernion’s explanation has been reported by Sina Finance and summarized by Ground News, but the available reporting does not include a formal Anthropic blog post or corporate statement specifically addressing the writing-style tradeoff the way the company did with its April 2026 Claude Code postmortem. No independently reproduced benchmark scores comparing Opus 4.6, Opus 5, and Opus 5.5 on human-readability metrics have been published alongside the claim, and Kernion’s quotes as reported come through syndicated and translated coverage rather than a verified verbatim English transcript from Anthropic itself. As of mid-October 2026, that gap in the public record still holds, with no follow-up Anthropic blog post or corporate statement specifically addressing the writing-style tradeoff surfacing beyond Kernion’s original comments.
That does not make the account unreliable, but it does mean the explanation should be read as an engineer’s candid, on-the-record account of an internal tradeoff rather than an audited corporate disclosure with attached data. Readers evaluating whether Opus 5.5 actually fixes the problem for their own use case are better served testing it directly than assuming the fix is complete based on Anthropic’s framing alone.
Frequently asked questions
Who explained why Claude’s writing got worse?
Jackson Kernion, an Anthropic engineer who works on Claude’s fine-tuning, publicly explained the writing-style shift on September 23, 2026, according to reporting from Sina Finance’s technology desk, summarized by Ground News.
Why did Claude’s writing get denser and more jargon-heavy?
According to Kernion, later Claude models were optimized more heavily for math, code, and technical explanations aimed at other AI systems, which shifted the model’s reward signals toward dense, machine-legible output at the expense of natural, human-readable prose.
Which Claude model was the last one with strong writing quality, according to Anthropic?
Kernion pointed to Claude Opus 4.6 as the last release he was fully satisfied with on writing quality, saying he hadn’t been as happy with a model’s writing since that release.
Does Claude Opus 5.5 fix the writing-quality problem?
Anthropic has framed Opus 5.5 as a targeted improvement to sentence clarity and information density, and Kernion described it as communicating more naturally. He also said the underlying tradeoff remains “a hard problem to solve,” so the fix appears partial rather than complete.
Is Opus 5.5’s writing as good as Opus 4.6’s now?
Not confirmed as a full return, according to Kernion. He named Opus 5.5 as the first release since Opus 4.6 that moves back toward that earlier standard, but stopped short of calling it equal to Opus 4.6, describing the underlying tension as “a hard problem to solve” that Anthropic will keep working on.
Does Opus 5.5 actually use fewer em dashes?
Yes. Anthropic says Opus 5.5 cuts em dash usage by 95% compared with Claude Opus 5, one of the few specific, measurable figures attached to the writing-clarity fixes Kernion described, alongside the broader goal of reducing dense, jargon-heavy “info dumps.”
Is this the same issue as the 2025 and 2026 Claude Code quality complaints?
No. The September 2025 and March-April 2026 Claude Code issues were traced to specific bugs and product-layer changes, including a reasoning-effort downgrade and a caching bug. Kernion’s September 2026 writing-quality explanation is about training-objective tradeoffs, a separate and distinct cause.
Did Anthropic intentionally degrade Claude to save money?
Anthropic has repeatedly denied intentionally “nerfing” its models to manage demand or cut costs, including in its April 2026 postmortem on Claude Code. Kernion’s September 2026 explanation frames the writing-style shift as an unintended side effect of capability optimization, not a deliberate cost-saving measure.
Are other AI labs facing the same writing-quality tradeoff?
The underlying mechanism, reward signals favoring machine-legible technical output over human readability, is common across labs training on similar reinforcement learning techniques. No comparable public, named-engineer explanation has been reported from OpenAI or Google addressing the same tradeoff in their own models.
What should developers do if they rely on Claude for human-facing writing?
Testing Opus 5.5 directly against specific writing tasks, rather than assuming the fix is complete, is the most reliable way to evaluate whether the update addresses a given use case, since no independently reproduced human-readability benchmark has been published alongside Anthropic’s claims.
Related
- Claude Opus 5.5 Matches Fable 5.1 at 60% Less Cost [2026]
- GPT-6 Sol, Luna Undercut Claude Opus 5.5 by 50% [2026]
- Claude Code vs Cursor 2026
- ChatGPT vs Claude vs Gemini vs Grok: 20-Point IQ Gap [2026]
- Anthropic’s 3-Step AI Slowdown Meets New Model Bid [2026]
- Claude Fable 5.1 vs GPT-5.6 Sol vs GLM-5.3: 96% SWE [2026]
Related Coverage
- ChatGPT vs Claude vs Gemini vs Grok: 20-Point IQ Gap [2026]
- xAI Launches Grok 4.7 at $2/$6, Trails Claude by 7 Points [2026]
- Claude Opus 5.5 Matches Fable 5.1 at 60% Less Cost [2026]
- Qwen3.8-Omni-Flash vs Qwen 3.8 Max vs Claude Fable 5.1 [2026]
- Set Up Claude Fable 5.1 API: 12 Steps, 90 Min [2026]
