{"id":95067,"date":"2026-04-06T14:53:49","date_gmt":"2026-04-06T21:53:49","guid":{"rendered":"https:\/\/github.blog\/?p=95067"},"modified":"2026-04-06T14:53:51","modified_gmt":"2026-04-06T21:53:51","slug":"github-copilot-cli-combines-model-families-for-a-second-opinion","status":"publish","type":"post","link":"https:\/\/github.blog\/ai-and-ml\/github-copilot\/github-copilot-cli-combines-model-families-for-a-second-opinion\/","title":{"rendered":"GitHub Copilot CLI combines model families for a second opinion"},"content":{"rendered":"\n

When you ask a coding agent to build a data pipeline, it may not use the best structure. But what if the agent got a second opinion before it executed the plan?<\/p>\n\n\n\n

Today, in GitHub Copilot CLI<\/a>, we’re introducing Rubber Duck in experimental mode<\/strong>. Rubber Duck leverages a second model from a different AI family to act as an independent reviewer, assessing the agent’s plans and work at the moments where feedback matters most.<\/p>\n\n\n\n

To catch different kinds of errors, a different perspective matters. Our evaluations show that Claude Sonnet + Rubber Duck makes up 74.7% of the performance gap between Sonnet and Opus alone, achieving better results for tackling difficult multi-file and long-running tasks. Use \/experimental<\/code> in Copilot CLI to access Rubber Duck alongside our other experimental features.<\/p>\n\n\n\n

The problem: Confident mistakes can compound<\/h2>\n\n\n\n

Today’s coding agents follow a clear loop. First, the agent assesses the task, then drafts a plan, implements, tests, and iterates if necessary. It’s a powerful flow that works well, but it has blind spots. Any decision an agent makes early on, especially in the planning stage, is the foundation you’re building upon. Assumptions and inefficiencies become dependencies, and by the time you notice, you may have to fix more than just the small mistake at the start.<\/p>\n\n\n\n

Using self-reflection and having the agent review its own output before moving forward is a proven technique. However, a model reviewing its own work is still bounded by its own training biases: the same training data and techniques, the same blind spots.<\/strong><\/p>\n\n\n\n

Rubber Duck adds a second perspective<\/h2>\n\n\n\n

Rubber Duck is a focused review agent, powered by a model from a complementary family to your primary Copilot session. When you’ve selected a Claude model from the model picker to use as your orchestrator, Rubber Duck will be GPT-5.4. As we experiment with Rubber Duck, we are exploring other model families for the orchestrator and for the Rubber Duck. The job of Rubber Duck is to check the agent’s work and surface a short, focused list of high-value concerns: details that the primary agent may have missed, assumptions worth questioning, and edge cases to consider.<\/p>\n\n\n\n

When does the cross-family review help?<\/h3>\n\n\n\n

We evaluated Rubber Duck on SWE-Bench Pro<\/a>, a benchmark of large, difficult, real-world coding problems drawn from open-source repositories. Here’s what we found:<\/p>\n\n\n\n

Claude Sonnet 4.6 paired with Rubber Duck running GPT-5.4 achieved a resolution rate approaching Claude Opus 4.6 running alone, closing 74.7% of the performance gap between Sonnet and Opus.<\/p>\n\n\n\n

We noticed that Rubber Duck tends to help more with difficult problems, ones that span 3+ files and would normally take 70+ steps. On these problems, Sonnet + Rubber Duck scores 3.8% higher than the Sonnet baseline, and 4.8% higher on the hardest problems identified across three trials. Here are a few examples of what Rubber Duck finds:<\/p>\n\n\n\n