MiniMax M2.7 tops GPT-5.4, Sonnet

Fair benchmarks are mostly measuring prompt robustness, not real production ceilings. For tutoring and agentic workflows, the prompt is part of the model stack, not a cosmetic detail.

• –MiniMax M2.7 jumped to 85% with its tuned prompt, ahead of Sonnet, GPT-5.4, and Gemini in this setup.
• –The generic “coding partner” prompt punished cheaper models that need explicit structure, while premium models handled vaguer instructions better.
• –The kid-simulator and pedagogical judges exposed variance that a single aggregate score would hide, especially across different child personas.
• –The cost story is the real punchline: a $0.30/M-token model can win if you spend the time to shape its behavior.
• –If you ship AI tutors, copilots, or agents, prompt engineering is likely the first lever to optimize before paying for a more expensive model.