Claude Opus 4.8

The 122-comment USER layer (HackerNews + Lemmy, top-upvote slice shown) is dominated by industry meta-discussion rather than hands-on Opus 4.8 experience. Product-specific signal that does exist: (1) Capability loyalty — a $200/mo Claude subscriber who also spends up to $200/mo extra on API tokens says they 'find it really hard to justify using a less powerful model' for real work. (2) Discernibility skepticism — a top comment argues frontier models are 'mostly past the point of human ability to discern whether they are actually better or worse than predecessors' and suspects benchmark saturation; another says the only trustworthy benchmark is one you build yourself and keep secret. (3) Reliability anecdotes — one user reports the model 'seemed to lose the ability to count in order' mid-task; another's crossword eval shows it mixing up 4-down/5-down clues despite otherwise impressive output. (4) Pricing friction — a comment notes Claude subscription prices 'increased 50%'. (5) Provenance controversy — a thread where 'Opus 4.8' self-identifies is dismissed by commenters as likely 'a scam proxy service claiming to sell Opus 4.8 API access but routing the requests to Qwen'. (6) Heavy off-topic content: Kimi K3 hardware requirements (1.5TB VRAM), GRAM/latent-reasoning research, and Anthropic's accusations that DeepSeek/Moonshot/MiniMax distilled its models — with sarcastic pushback ('It's bad when China does it. But the other way around... that's just business'). Net: users respect the capability but distrust the pricing narrative, the benchmarks, and even some resellers' model identity.

No aggregate ratings were found for this product during the last harvest.