Observations on Opus model Role Token changes
Ratter · x · 2026-09-01
The author observes a change in Opus's training format, which used to be plain text \n\nHuman: / \n\nAssistant:. This seems to have changed around version 4.8.
Now the model can output \n\nHuman: without stopping generation, suggesting a shift in the underlying Role Token handling mechanism.
Related event: Users Say Opus Training Shift to XML Tags Explains Leak Behavior(3 posts)→
More from Models
- SparkLLM releases on-device models with 1M context support — MaziyarPanahi · 2026-09-01
- Opus 5 criticized for extreme verbosity and injecting irrelevant rants — haider1 · 2026-09-01
- SparkLLM open-sources X2.5 4B/1.7B on-device models with native 1M-token context — vllm_project · 2026-09-01
- OpenAI Codex sees exponential growth amid Anthropic user decline — FinanceYF5 · 2026-09-01
- Hypothesis on Opus 5 leakiness: Mixed old and new training formats — Ratter · 2026-09-01
- Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored Trends on HF — DavidAU · 2026-09-01