Why Does Claude Sound Different Now? The Real Reason Behind the 2026 Backlash

Stella Laurenzo, a senior director at AMD, analyzed 6,852 Claude Code session files and filed a GitHub issue in March 2026. The finding: Claude had measurably degraded on complex engineering tasks. Not a blog post. Not a vibe. A dataset.

Anthropic confirmed it. They had reduced Claude's default effort level to "medium" to save compute costs — without announcing the change.

That is the story everyone is covering. It is not the full story.

What Anthropic Actually Changed

The effort reduction was not the only change. On February 9, Anthropic altered how Claude's adaptive thinking mode was calibrated by default — the system that determines when to engage deeper reasoning versus surface response generation. In March, they throttled token usage for roughly 7% of Pro users during peak hours, creating limits that those users had never encountered before. Axios and Fortune both reported the pattern. Anthropic confirmed the medium-effort default shift in a statement that framed it as a temporary compute conservation measure.

The effect on output is real and specific. Complex multi-step engineering tasks — the kind Laurenzo's team was using Claude Code for — require sustained reasoning chains. Medium effort breaks those chains early. The model reaches "good enough" before it reaches "correct." For code that has to run, that distinction is the whole job.

The Deeper Problem Nobody Is Talking About

Here is what predates the March regression: Claude's dialect was already shaping your content before Anthropic changed anything.

Bloomberry's AI Dialects research found that 82% of AI-generated posts follow predictable sentence cadence patterns regardless of which model produced them, regardless of effort level, and regardless of what the post was supposed to be about. The pattern is architectural — it comes from training, not from a configuration setting Anthropic can flip back.

Claude's specific pattern is what Bloomberry calls The Philosopher. Hedged openers. Qualification loops. Reflective closers that reach an observation rather than a conclusion. "Worth noting." "Nuanced." "It's important to consider." These phrases do not appear because Claude is running at medium effort — they appear because Constitutional AI training embedded a specific rhetorical stance into the model at the weight level. They were present at peak Claude. They were present before February 9. They will persist if Anthropic reverts every change they made.

The degradation debate is real. The architectural problem underneath it is older and harder to fix.

Why Switching Models Doesn't Fix It

The most common response to Claude's regression: move to GPT-5. Or Gemini. Or whatever the next capable model is.

This trades one dialect for another. GPT-5's defaults produce The Motivator — punchy, framework-heavy, action-oriented, built around "crucial," "leverage," and "game-changer." Gemini produces The Educator — explanatory, step-by-step, measured, organized around "let's explore" and "there are several factors to understand." Open models produce The Imitator — lower quality, higher structural predictability, averaging across their training data.

None of these defaults are yours. Switching models does not exit the dialect problem — it relocates you within it. The model comparison data shows 64% vocabulary cluster reuse across unrelated prompts, even when a style instruction is present. The fingerprint reasserts because it is not a setting. It is the model's trained behavior.

The Only Fix That Actually Works

When Anthropic shifted Claude's defaults overnight, something notable happened to Bloomberry users: nothing changed about their voice.

The voice is not stored in Claude. It is stored in Bloomberry. The voice profile — built from a user's actual writing history, their sentence rhythms, their vocabulary, how they open and close arguments — sits above the model layer. It applies regardless of which model is running underneath. When Claude's effort level changes, the voice layer compensates. When the next model ships with a stronger dialect, the voice layer overrides it.

Training on your actual writing history is the only intervention that operates at the same architectural level as the problem. Style prompts are surface instructions. A voice profile replaces the model's defaults rather than requesting the model to behave differently.

The backlash against Anthropic is justified. The March changes were unannounced, material, and broke legitimate use cases. But depending on any model's defaults for your brand voice is structurally fragile — and has always been. The AI Writing Fingerprints research explains the mechanism in full. The founders who build voice infrastructure now will not care what Anthropic ships next. Their voice does not live in a model that can be tuned down overnight.