MiniMax M2.7 plugged into Claude Code via API and tested against Claude Opus 4.7 on PyTorch refactoring, Obsidian note drafting, and a Kaggle competition submission.
Key Takeaways
M2.7 routed through Claude Code via ANTHROPIC_BASE_URL override; all four model env vars pointed to MiniMax-M2.7 on MiniMax’s $40/month Plus plan.
Cost roughly $8 equivalent at PAYG rates ($0.30/$1.20 per million input/output) across 91M tokens over five days; comparable Opus 4.7 usage would run ~10x higher.
Speed was subjectively ~2x faster than Opus 4.7 on multi-step agentic tool calls; no rate-limit hits across intensive sessions.
Both M2.7 and Opus 4.7 failed identically on Kaggle kernel-only mechanics (leaking the exposed target column into features) until the constraint was stated explicitly in the prompt.
Prompt-bootstrapping trick: ask M2.7 to read its own output alongside Opus output, then generate an improved prompt tuned for itself before the next iteration.
