MiniMax M2.7: Self-Evolving AI Model Matches Sonnet 4.6 at 1/3 Cost — 56.22% SWE-Pro, 97% Skill Adherence

CnTechPost / Latent Space AINews·high signal

MiniMax released M2.7 on March 18, 2026, a model that participates in its own training loop — handling 30-50% of its R&D workflow autonomously. It scores 56.22% on SWE-Pro and 55.6% on VIBE-Pro, matching Sonnet 4.6 benchmarks at $0.30/$1.20 per 1M input/output tokens, roughly one-third the cost of GLM-5. Available immediately on MiniMax Agent, API, Ollama, OpenRouter, and Vercel.