Performance difference between Github Copilot Premium vs API models

I've been using Claude Sonnet 4 for coding in VS Code, and I'm noticing a significant difference in performance between the premium model (bundled with subscription) and the same model accessed via my Anthropic API key.

What I'm experiencing:

• API version: Follows instructions precisely, handles rambling/poorly formatted requests effortlessly, excellent at complex coding tasks
• Premium bundled version: Good but inconsistent, sometimes misses the mark, occasionally breaks existing code

I've tested with identical prompts and the difference is consistent - the API version just "gets it" while the premium version sometimes struggles with the same requests.

The problem: I've burned through my API credits and need to rely on the premium version, but the performance gap is frustrating.

Questions for the community:

1. Has anyone else noticed this difference?
2. Are there specific prompting techniques that work better with the premium version?
3. Any settings or approaches to make premium Claude perform closer to the API version?