MAI Code 2 hands-on: replacing Claude as orchestrator, tops Terminal-Bench

imdigitalashish · x · 2026-11-21

The author shares a hands-on review of Microsoft's MAI Code 2, calling it a big leap. His workflow: MAI Code 2 as the main agent with gpt-6.2 and MAI Code Flash as subsessions, occasionally asking Claude as a subagent for cross-checks — which mostly agreed with MAI Code 2.\n\nHis previous pain point with Claude + GPT was Claude's inexplicable moral lectures and refusals (e.g., refusing to use Playwright to backfill CUJO mappings); MAI Code 2 fixed that. He argues criticism of GPT-6.2 mirrors what happened to Claude Opus 5: not fun to talk to, but better with fewer mistakes when driven by another model. He also notes one wrong Copilot Fusion auto-router decision can drift the whole task and is hard to steer back, so he avoids Fusion while still investigating. MAI Code 2 ranks #1 on Terminal-Bench and sits on the Pareto frontier for both image and harness.

Original post →

More from coding & agent

coding & agent channel →