# Theory: Opus 5’s bizarre communication style was an anti-distillation feature, not a bug — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/1b73ab0b04dcc835

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-09-27 00:31 ingested 2026-09-28 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1wr64ez/theory_opus_5s_bizarre_communication_style_was_an/)

## retained text



```
I have a theory that Opus 5’s strangely poor communication style was an anti-distillation strategy, a kind of Trojan horse for anyone training on its outputs. 

The idea is that a company using Opus 5’s outputs as training data would end up sabotaging its own models, even though Opus 5 itself would still perform well on coding tasks.

The unfortunate side effect is that developers who actually read and interact with the model’s output while coding—rather than just letting it work autonomously and judging the resulting code solely based on its performance—had to deal with Opus 5’s bizarre verbal style and confusing explanations. That made the model extremely frustrating to use for many people despite its underlying technical capabilities.

Opus 5.5, which is much more pleasant and sensible to interact with (for now at least), happens to also come with a brand new anti-distillation mechanism called “Preserved Thinking.” This protects the model’s reasoning history by preserving conversation state and preventing API users from retroactively changing system prompts around preserved reasoning blocks.

If Anthropic has now implemented a more direct mechanism for blocking reasoning-focused distillation, perhaps it no longer needs to rely as heavily on making the model’s visible outputs poor training data.

This would provide a possible explanation for what I otherwise find pretty incomprehensible about Opus 5: why would Anthropic intentionally release a model that was so obviously difficult to understand and articulate decisions with? It strains credulity to imagine that the Anthropic team wasn’t aware of this issue when they released the model. If its corrupted communication style was serving some other purpose, however, the decision starts to make sense.

I’m curious what others think. Is this technically plausible as an anti-distillation strategy, or is there a better explanation for why Anthropic knowingly shipped Opus 5 with such a frustratingly poor and ineffective communication style?
```



document fetched 2026-09-28 00:24 · canonical https://reddit.com/r/ClaudeAI/comments/1wr64ez/theory_opus_5s_bizarre_communication_style_was_an
