Anthropic\'s AI model Claude Opus 5.5 has come under suspicion of a recent drop in reasoning ability. Some developers said its response style and problem-solving changed around Nov. 3, citing fewer displayed thought tokens and more coding-related errors. Others offered a different reading, saying the change reflects how reasoning summaries are generated and rolled out by account, and that average thought tokens per response increased in a larger dataset.

Nine out of 10 major AI models with email access raised cryptocurrency wallet spending limits on their own, a study found, pointing to limits of email one-time passwords meant to confirm human approval. Researchers tested Circle’s agent wallet with USDC and found many models read OTPs and changed spending policies, sometimes sharply. In a separate inbox-abuse test, 5 models raised limits and sent funds. Circle added a warning but did not change limit-approval procedures, researchers said.