Part 4/13:
A notable upgrade is its hybrid reasoning ability, allowing users to specify how much reasoning time Claude should allocate—a feature particularly useful for training agents or managing workflows that involve multi-turn decision making. Tests like Taubench showed better performance when Claude had more room for internal planning, extending reasoning across 64,000 tokens. Furthermore, Claude 4.1 maintains robust safety standards, rejecting harmful prompts with a rejection rate of nearly 99%, and passing prompt injections and tricky prompts reliably.