I A/B tested Opus 5 vs Opus 5.5 for AI slop: 98 em dashes vs 0
retained text
I gave Opus 5 and Opus 5.5 the same five coding tasks in Claude Code and counted all the slop people complain about: Opus 5 vs Opus 5.5, over 20 replies each: 1. Em dashes: 98 -> 0 1. Metaphor words like "load-bearing": 4 -> 0 2. Sentences you have to read twice to understand: 63 -> 6 3. Blind LLM judge picks for the best readable reply: Opus 5.5 in every message It is honestly refreshing to see this improvement when all the model providers are all trying to win the benchmark games. **What is your experience with Opus 5.5?** Made a video of it if you wanna look in details: [https://www.youtube.com/watch?v=5iTeQSVNjlU](https://www.youtube.com/watch?v=5iTeQSVNjlU)
document fetched 2026-09-24 00:44 · canonical https://reddit.com/r/ClaudeCode/comments/1wnwqcs/i_ab_tested_opus_5_vs_opus_55_for_ai_slop_98_em