A quick review on @MiniMax_AI M3: Harness: Roo Code within Windsurf Environment - A very capable model for coding. I tested using Code mode with a medium size task that goes across multiple components (complex graph on canvas + inspector for scenario mode). It was able to execute quite flawlessly in one shot (800+ lines of code) and also ensure the design matches existing UI. - The most surprising part was planning. It's not just efficient. I tested in Architect mode without a complete instructions on what components should be touched and it went ahead and found everything related to the task by itself. This is something Claude Opus 4.8 could never do (coz it doesn't fully listen nor complete task) and in this case, even GPT 5.5 doesn't get to if I don't provide extensive instructions. - Total cost for the task: $0.15 or around 20-30x cheaper in comparison to GPT 5.5 and Claude Opus 4.8. Verdict: M3 got close enough for me to replace specific tasks I would normally use GPT 5.5 for and the team will start using it as backup option when they go over the limit on Codex and Claude Code. There were still cases I had to switch back to GPT 5.5 to complete, but I expect 50%+ of the tasks can now be done using M3 instead. I will continue to work with Minimax team with the goal to completely replace GPT 5.5 from our workflow and use future GPT models for much larger and complex tasks instead.