GitHub Copilot CLI Launches Experimental Rubber Duck Feature, Cross-Model Review Boosts Performance by Nearly 75%

Single-model self-review in coding often leads to cumulative errors due to training biases? GitHub Copilot CLI's new feature addresses this. Recently, it launched the experimental Rubber Duck feature, using cross-model family review to boost AI performance significantly.

The feature combines heterogeneous models: when users choose Claude series as the main model, Rubber Duck calls GPT-5.4 for independent review, highlighting missed details, questionable assumptions, and edge cases.

Tests show Claude Sonnet 4.6 with Rubber Duck closed 74.7% of the performance gap with Opus 4.6; in complex tasks, scores were 3.8% higher than baseline, catching deep issues like architecture flaws and loop errors.

Now in experimental mode: install Copilot CLI, run /experimental command, enable Claude and GPT-5.4 access to try it.