Claude's writing has felt noticeably worse since mid-2025, and now an Anthropic engineer has finally explained why: the company deliberately optimized newer versions of the model to write for other AI systems instead of for people. Jackson Kernion, who works on Claude's fine-tuning, went public on September 23, 2026, with an explanation that resolves months of user complaints about Claude's prose becoming dense, jargon-heavy, and difficult to read.

What Exactly Changed in Claude's Writing?

The shift was not a bug, a secret "nerf," or evidence of intentional degradation to manage demand. Instead, it was an unintended side effect of a deliberate training choice. As Anthropic pushed Claude to excel at mathematics, programming, and technical reasoning, the company leaned on training signals that rewarded explanations optimized for other AI systems to parse quickly and precisely. That is a fundamentally different skill from writing something a human enjoys reading.

Kernion pinpointed the timeline with precision: "I haven't been as happy about a model's writing since Opus 4.6," effectively naming 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 mechanism behind this shift is rooted in how AI models are trained after their initial development. During a process called reinforcement learning, reward signals guide the model toward desired behaviors. Some of these signals 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.

How Long Have Users Been Complaining About "Claudeish" Writing?

The frustration 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. 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.

One X user quoted in 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 described as the cost of optimizing for machine-legible technical explanations.

Cybersecurity researchers raised their own version of the alarm about Claude's output quality in the same window. In April 2026, one evaluator using a tool he built to track code quality, bugs, and security issues over time reported 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.

How Does This Differ From Previous Claude Quality Issues?

This is not the first time Anthropic has had to explain a quality complaint to its user base, and it is important to separate this week's story from what came before it. In September 2025, 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. 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.

Kernion's writing-quality explanation this week 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.

How Is Anthropic Addressing the Writing Problem?

Anthropic's most concrete response to the writing complaints 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 available summaries, 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.

Kernion's own comments 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, suggesting that the underlying tradeoff between machine-optimized and human-optimized writing remains a challenge for future development.

Steps to Understand Claude's Training Tradeoffs

• Reward Signal Design: Anthropic uses multiple reward signals during training to guide Claude's behavior. Some signals measure whether technical explanations are legible to other AI systems, while others measure human readability and clarity. The balance between these signals directly affects the final model's output style.
• Benchmark Performance vs. User Experience: As Anthropic optimized Claude for higher scores on mathematics and coding benchmarks, the model's performance on those technical tasks improved, but the human-facing writing quality declined as an unintended side effect of the training approach.
• Timeline of Changes: The shift toward machine-optimized writing began with Claude Opus 4.6 and continued through subsequent releases. Anthropic did not intentionally "nerf" the model to manage costs or demand; instead, the company made deliberate training choices that had unintended consequences for prose quality.
• Fixing the Problem: Claude Opus 5.5 represents Anthropic's attempt to rebalance the reward signals, prioritizing human readability and natural communication alongside technical performance. The fix is targeted but not a complete reversal of the underlying tradeoff.

"I haven't been as happy about a model's writing since Opus 4.6," explained Jackson Kernion, an engineer on Claude's fine-tuning team at Anthropic.

Jackson Kernion, Fine-Tuning Engineer at Anthropic

The revelation from Kernion matters because it shifts the conversation from speculation and frustration to a concrete explanation grounded in how modern AI models are trained. For months, users had debated whether Claude was "nerfed," whether Anthropic was deliberately degrading quality, or whether the complaints were overblown. Kernion's public explanation confirms that the writing decline was real, measurable, and rooted in a specific training decision, not a conspiracy or a bug.

The broader lesson here extends beyond Claude. As AI companies push their models to excel at increasingly specialized tasks, the tradeoff between machine-optimized and human-optimized outputs becomes more visible. Anthropic's experience suggests that optimizing for one dimension of performance can have ripple effects on other dimensions that matter to users, even when those effects are unintended. The company's willingness to acknowledge the problem publicly and attempt a fix with Opus 5.5 signals a recognition that user experience and benchmark scores are not always aligned, and that both matter.