Descriptions:
Sam Witteveen provides a detailed breakdown of MiniMax M2.5, a frontier-competitive large language model from one of China’s leading AI labs that costs approximately $1 per hour of continuous operation — compared to $15–20 per hour for Claude Opus. The model ships in two API variants: a standard 50 tokens/second version and a Lightning variant at 100 tokens/second, priced at $0.30 per million input tokens and $2.40 per million output tokens.
The analytical centerpiece is a benchmark blog post from OpenHands (formerly OpenDevin, originating from Carnegie Mellon), which tested M2.5 as the top open model for software engineering tasks — finding it more than 90% cheaper than Claude Opus for comparable coding workloads, though still trailing GPT-5.2 Codex at the high end. Witteveen also highlights that Ollama has partnered with MiniMax to serve the model via their cloud tier, and notes the weights have been shared with multiple serving providers despite not being fully open yet.
On the architecture side, MiniMax credits its rapid improvement trajectory — three M2-generation model releases in 108 days, faster than Anthropic, Google, or OpenAI — to aggressive reinforcement learning scaling via their Forge framework, trained across hundreds of thousands of internal task environments derived from real company workflows. Witteveen closes by framing M2.5 as a potentially compelling default for always-on agents like OpenClaw, where the cost differential between $1/hour and $20/hour compounds dramatically at sustained runtime.
📺 Source: Sam Witteveen · Published February 13, 2026
🏷️ Format: Review
