MiniMax M2.5 Promises Intelligence Pricing Too Low to Measure

Shanghai-based AI firm MiniMax has officially unveiled its latest open-weight model, M2.5, released under the MIT license.

According to MiniMax, M2.5 was trained via reinforcement learning across hundreds of thousands of complex environments. This methodology enables robust performance with lower token consumption, particularly excelling in long-horizon tasks.

The model features autonomous planning capabilities to optimize actions, distinguishing it from models that only plan when prompted. The company also highlights its ability to process Word, Excel, and PowerPoint documents rapidly and smoothly.

Published benchmark results indicate top-tier performance in coding: 80.2% on SWE-Bench Verified, 79.7% on the Droid platform, and 76.1% on OpenCode. In certain scenarios, it even surpasses Anthropic's Claude Opus 4.6.

MiniMax claims M2.5 outperforms GPT-5.2, Gemini 3 Pro, and Claude in web browsing (BrowseComp: 76.3%), AI agent tool usage (BFCL: 76.8%), and office tasks involving Excel.

Cost-Effective and High-Speed

MiniMax offers two variants, including the M2.5-Lightning, which processes 100 tokens per second—twice the speed of other leading models. Data suggests continuous operation costs just one dollar per hour. This pricing makes the model ten to twenty times cheaper than comparable offerings from Anthropic, Google, and OpenAI. The company states that four M2.5 instances can run continuously for a full year at a cost of $10,000.

M2.5 is available for free on Hugging Face and GitHub. A detailed technical report will be published on minimax.io. API access and dedicated coding plans can be found at platform.minimax.io.

The company stated, "M2.5 is the first frontier model where users need not worry about costs, fulfilling the promise of affordable, immeasurable intelligence." Such pricing could impose significant pressure on Western AI labs, which are growing rapidly but remain unprofitable.

Reports indicate that some US startups are switching to Chinese models due to lower prices, although the US enterprise AI market remains dominated by Microsoft, Google, OpenAI, and Anthropic. Even Chinese labs acknowledge that US counterparts maintain a lead, exacerbated by chip scarcity. In other words, the situation is not black and white.