Claude Opus 5.5 uses 95% fewer em dashes, but its answers are getting

Anthropic's Claude Opus 5.5 is writing differently than its predecessor, and the shift is measurable. New analysis published via BleepingComputer shows the model has cut its use of em dashes by roughly 95 percent compared with Claude Opus 5, while also producing shorter sentences and simpler word choices — changes that make its output look less like stereotypical AI prose.

This piece breaks down what the writing-style shift means for developers and teams shipping Claude Opus 5.5 in production, how to pull the update, and where the tradeoffs land. If your pipelines depend on Claude's output format — or if you've spent time post-processing AI text to remove telltale punctuation — this update changes the calculus.

What Claude Opus 5.5 uses 95% fewer em dashes shipped

Anthropic did not announce a dedicated punctuation or style overhaul; the behavioral changes emerged from independent analysis comparing Claude Opus 5.5 output with Claude Opus 5 across matched prompts. Researchers found a 95 percent reduction in em dash usage, alongside measurable drops in average sentence length and vocabulary complexity. The analysis did not find that the model became less capable — benchmarks and task performance held — but the stylistic fingerprint changed substantially.

The timing is notable. Anthropic has been under pressure from enterprise customers and researchers who use AI-detection tools as part of editorial or compliance workflows. A model that writes with fewer recognizable AI-writing tics is more useful in contexts where blending with human prose matters. Anthropic has not publicly confirmed the style changes were deliberate, but the consistency across prompts suggests a systematic shift rather than noise.

What changed for builders in Claude Opus 5.5 uses 95% fewer em dashes

The concrete difference is that text produced by Claude Opus 5.5 is shorter per sentence, uses simpler vocabulary, and almost never reaches for the em dash that has become a widely cited AI-writing marker. For teams running Claude in document drafting, customer-facing copy, or any workflow where text is published without heavy human editing, that matters immediately.

One practical implication: post-processing scripts that strip or replace em dashes before publishing can be relaxed or removed. The flip side is that some teams report the model's answers are getting longer in total word count even as individual sentences shorten — more paragraphs rather than denser sentences. If you have hard output-length constraints in your prompts, verify them against Opus 5.5 before cutting over.

How to install or upgrade Claude Opus 5.5 uses 95% fewer em dashes

If you're running Claude Code, update the CLI to pick up the latest model routing:

claude update

To switch the active model inside a Claude Code session:

/model claude-opus-5-5

To launch a session pinned to Claude Opus 5.5 from the command line:

claude --model claude-opus-5-5

To set Claude Opus 5.5 as your default model so every new session uses it automatically:

claude config set model claude-opus-5-5

For teams calling the API directly, update your model parameter in request payloads to claude-opus-5-5 and redeploy. No SDK version bump is required beyond having a current anthropic package; the model identifier is the only change needed at the call site.

Gotchas and compatibility in Claude Opus 5.5 uses 95% fewer em dashes

The most significant compatibility concern is output length. Because Claude Opus 5.5 favors shorter sentences but may produce more of them, total response length can increase even when you haven't changed your prompts. Teams with token-budget constraints or UIs that render responses inline should run a comparison batch before switching production traffic.

Style-dependent evals are another risk area. If your automated evaluation suite scores responses against a Claude Opus 5 baseline — checking for phrasing patterns, punctuation density, or sentence rhythm — scores may drop not because quality fell but because the distribution shifted. Updating your golden-set examples to reflect Opus 5.5 output before running regressions will save confusion. There is no indication from the available analysis that factual accuracy, instruction-following, or coding performance regressed; the changes appear isolated to writing style.

What to watch after Claude Opus 5.5 uses 95% fewer em dashes

The writing-style shift raises an open question about whether Anthropic will extend similar changes to Claude's smaller models. Haiku and Sonnet variants still carry the older stylistic patterns to varying degrees, so teams using a tiered routing strategy — Opus for complex tasks, Sonnet or Haiku for high-volume cheaper calls — may see inconsistent prose style across tiers. Anthropic has not announced a timeline for aligning the full model family.

Longer term, the analysis that surfaced these changes was independent, not from Anthropic's own model cards or release notes. As AI-detection tooling grows more sophisticated and as enterprise style guides evolve to account for AI-assisted writing, the pressure on model providers to make style behavior legible and controllable will increase. Watch for Anthropic to either formalize style parameters — giving developers direct control over punctuation density and sentence cadence — or to document these behavioral changes explicitly in future model cards rather than leaving them to third-party reverse engineering.