Opus 5 is a very interesting release: benchmarks are useless, new training method, declining pleasantness, and are we failing at alignment?

The author analyzes Opus 5: general benchmarks are practically useless; Anthropic tried a new training approach (train Mythos first, then distill) but results are mixed; models are becoming less pleasant to work with, and AI might be steering humans – suggesting alignment failure.

• Anthropic
• training
• alignment
• benchmarks
• Opus 5