{"slug": "get-ready-for-fewer-opus-5-5-em-dashes", "title": "Get ready for fewer Opus 5.5 em-dashes", "summary": "Data from the agent-duel leaderboard, a battle system built for OpenCode, confirms a Hacker News claim that Anthropic's Opus 5.5 was weakened, showing the model's reasoning output fell 50.6% relative to matched opponents, with a bootstrap breakpoint test across 16 candidate dates pointing to Nov. 1 (p-value 0.00320) rather than the Nov. 3 date claimed in the post. The same data shows Opus 5.5's em-dash frequency dropped from 14.76 to 8.40 per unit of text while opponents' barely moved (7.01 to 6.76, adjusted p-value 0.0440), and responses starting with \"I'm\" rose from 20.8% to 38.9% (adjusted p-value 0.0170). The author speculates the changes reflect a new model being tested rather than a simple nerf, noting em-dashes are a major user complaint about Opus.", "body_md": "# Get ready for fewer Opus 5.5 em-dashes\n\nA few days ago, a popular [Hacker News post](https://news.ycombinator.com/item?id=50422069) argued that Opus 5.5 had been nerfed, showing that reasoning tokens declined and language use changed.\n\nIn the thread, someone asked if anyone had proof of this, and I realised we did! We are working on a [battle system for OpenCode](https://github.com/bottomless/agent-duel), which gives us [data](https://agentduel.com/leaderboard) of how Opus 5.5 behaves in comparison to opponents.\n\n## Confirming the post\n\nFirst, we checked if the post is confirmed by our data, and it is! Opus’s reasoning output fell 50.6% relative to matched opponents!\n\n## Confirming the date\n\nThe post claims Opus was nerfed on Nov. 3, but looking at a chart of our data, this did not seem to be the case. In fact, Nov. 1 looked more likely.\n\nTo investigate this timing, we tested 16 candidates dates with a bootstrap breakpoint test. Nov. 1 produced the strongest reasoning-ratio break, although Nov. 2 and Nov. 3 could easily be the break date, too.\n\n| Candidate cutoff | Relative change | p-value | \n|---|---|---|\n| October 29 | −36.4% | 0.11351 | \n| October 30 | −36.4% | 0.13598 | \n| October 31 | −40.5% | 0.06444 | \n| **November 1** | **−50.6%** | **0.00320** | \n| November 2 | −47.3% | 0.00353 | \n| November 3 | −45.2% | 0.00347 | \n\n## Looking at language changes\n\nThe post also claims language changes in early november, so we looked into this too.\n\nAll three measures moved in the reported direction: more “I’m,” less “I should” or “I need to,” and fewer dashes. Two passed correction for multiple testing.\n\n| Wording feature | Before → after | Adjusted p-value | \n|---|---|---|\n| Starts with “I’m” | 20.8% → 38.9% | 0.0170 | \n| Uses “I should” / “I need to” | 57.0% → 43.8% | 0.0850 | \n| Contains em dash or `--` | 48.0% → 28.9% | 0.0000572 | \n\n## Further investigations – the em-dashes\n\nI was curious if these words changed, what else might have changed, and I found something interesting! Opus’s adjusted em-dash frequency fell about 43%, while its opponents’ barely changed.\n\nTo find candidate words, we ranked words, punctuation, and whitespace by their overall frequency across both periods, then ran a statistical test. (To be honest, I was worried this was a fishing expedition so a normal p-val wouldn’t work and Astra suggested a BH-25 p-val).\n\nFive patterns had adjusted p-values below 0.05:\n\n| Pattern | Opus before → after | Opponents before → after | Adjusted p-value | \n|---|---|---|---|\n| in | 77.18 → 99.89 | 77.49 → 79.84 | 0.0207 | \n| I’m | 90.58 → 155.07 | 46.96 → 54.25 | 0.0324 | \n| **Em dash —** | **14.76 → 8.40** | **7.01 → 6.76** | **0.0440** | \n| Hyphen - | 37.14 → 31.51 | 42.48 → 45.94 | 0.0440 | \n| the | 611.87 → 780.69 | 528.20 → 526.03 | 0.0462 | \n\n## Speculation\n\nSince em-dashes are a major complaint about Opus, it seems this goes beyond nerfing and is actually a new model being tested.\n\nThe increase in “I’m” is also interesting. Any system has a set of affordances that help a user interact with it, and to some extend personhood is the affordance of an AI agent. It seems Opus is going to lean even harder in this direction.", "url": "https://wpnews.pro/news/get-ready-for-fewer-opus-5-5-em-dashes", "canonical_source": "https://micjm.com/blog/get-ready-for-fewer-opus-5.5-em-dashes/", "published_at": "2026-11-11 23:40:32+00:00", "updated_at": "2026-11-11 23:50:22.556480+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-research"], "entities": ["Anthropic", "Opus 5.5", "Hacker News", "OpenCode", "agent-duel"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/get-ready-for-fewer-opus-5-5-em-dashes", "markdown": "https://wpnews.pro/news/get-ready-for-fewer-opus-5-5-em-dashes.md", "text": "https://wpnews.pro/news/get-ready-for-fewer-opus-5-5-em-dashes.txt", "jsonld": "https://wpnews.pro/news/get-ready-for-fewer-opus-5-5-em-dashes.jsonld"}}