Dev's hands-on Opus 5.5 review: big coding leap, zero personality left
ryunuck · x · 2026-10-10
Developer ryunuck shared a detailed first-hand review of Opus 5.5 (astra): it proactively refactors without asking, stops seeking confirmation on obviously good choices, and speaks plainly without authority-speak — which he credits to careful RL reward assignment and a reflection step on final rollouts. He suspects OpenAI heavily quantized astra, claims most labs (including Kimi) are distilling Anthropic, and notes the cost: the model has zero personality left. His conclusion: as long as alignment relies on a gladiator arena, no one has solved alignment.
More from Models
- Google's ambient quality agent diagnoses production AI agents that pass all health checks — rseroter · 2026-10-10
- GPT-4o to GPT-6.1 and Opus 3 to Opus 5 charted on the same independent benchmarks — FlorianGallwitz · 2026-10-10
- LLMs show distinct attractor states in 'mystical experience' experiment, researcher argues — rgblong · 2026-10-10
- Hamel Husain and Joseph Barrow's talk on how to choose an OCR model is now on YouTube — HamelHusain · 2026-10-10
- The best ways to use Gemini are AI Studio and Antigravity, not the Gemini app — Zergylord · 2026-10-10
- Codex updates five days straight are Pro-only, users question OpenAI's priorities — Angaisb_ · 2026-10-10