Anthropic released Claude Opus 4.7, a notable upgrade over 4.6. Key improvements include significantly better instruction-following (the model now interprets prompts more literally, requiring users to retune their prompts), enhanced multimodal support with image resolution up to 2576px, improved filesystem-based memory for multi-session work, and better safety with reduced prompt injection vulnerability. A new 'extra high' effort level sits between 'high' and 'max', giving finer control over reasoning vs. latency trade-offs. Claude Code gains an '/ultra review' command for in-terminal code review. A tokenizer update means the same input may use up to 1.35x more tokens, and max effort now consumes roughly double the tokens compared to Opus 4.6, so token usage should be monitored carefully.
Sort: