MiniMax M2.7: Self-Evolving Model Matches Sonnet 4.6 at One-Third the Cost

MiniMax shipped M2.7 on March 18 and made a claim nobody else has made with receipts: the model participated in its own R&D cycle. Not "we used AI to help train it" marketing. MiniMax says M2.7 autonomously handled 30–50% of the development workflow — reading logs, debugging failures, analyzing metrics, and optimizing performance scaffolds — while humans focused on architecture decisions and safety. The result: a model that scored 56.22% on SWE-Pro (matching GPT-5.3-Codex), 55.6% on VIBE-Pro (near Opus 4.6 parity), and hit a 97% skill adherence rate across 40+ complex skills exceeding 2,000 tokens each. The GDPval-AA ELO of 1495 is the highest among open-source models. CnTechPost Latent Space MiniMax

The pricing is where this gets real. At $0.30 input / $1.20 output per million tokens, M2.7 costs roughly one-third of GLM-5 while matching its reasoning benchmarks. Production incident recovery times dropped to under three minutes in real-world engineering scenarios. The model achieved a 66.6% medal rate across 22 ML competitions — the kind of metric that matters because competition submissions are adversarial by nature.

Available today on MiniMax Agent, their API, Ollama, OpenRouter, and Vercel. If you're building agent pipelines and paying frontier-model prices for Sonnet-class reasoning, M2.7 is the first credible alternative where the benchmarks, the price, and the availability all line up simultaneously. The self-evolution angle is the longer-term story — if models can meaningfully participate in their own improvement loops, the gap between releases compresses.