Your Opus 4.7 selection is working. Credits are correctly being deducted for Opus, and the main reasoning work uses that model.
What you’re seeing is the orchestrator pattern: Cursor sometimes uses a faster model (Composer 2.5 Fast) as the coordinator that delegates heavier work to Opus 4.7 via subagents. When you ask “what model are you?” and the coordinator responds, it honestly says Composer. When the Opus agent responds directly (as in your second screenshot), it says Opus. Both are truthful.
To force everything to use your selected model, enable Max Mode in the chat settings (the toggle next to the model picker).
This has come up before - more details in this thread and this one where staff confirmed that a global setting for subagent model selection is coming soon.