A developer spent three days running Claude Opus 4.7 and 4.6 side by side in real coding sessions. The verdict: the newer model is worse at getting things right on the first try, costs more, and uses fewer tools per turn.
4.7 prefers to reason internally rather than read project files. That leads to hallucinated project structures. It invents file paths and module relationships that don't exist, then writes code around those fictions.
The tool usage shift is the tell. 4.6 would read files directly, building understanding from what's there. 4.7 makes fewer tool calls per turn, trying to work things out in its head first. Sounds more sophisticated. In practice it means the model guesses more than it checks.
One Hacker News commenter named alegd, who uses Claude Code daily, said 4.7 "feels different" but couldn't pin down exactly how. Whether better context through files like CLAUDE.md would fix the accuracy gap is an open question.
Higher cost plus lower first-try accuracy is a rough combo for anyone paying by the token.