Quality per Dollar Squeezes MiniMax

DeepSeek changed the basis of competition from model quality alone to model quality per dollar, and that is brutal for standalone labs like MiniMax. When one rival ships a strong reasoning model at much lower API prices and with open weights, everyone else has to respond by charging less, giving away more, or both. That squeezes gross margin right when training and inference bills are still enormous.

• MiniMax is already selling its core text models cheaply, with M2.5 priced at $0.15 per million input tokens and $1.20 per million output tokens, and it pitches that as far below Claude level task costs. That shows the company is competing on economics, not just raw capability.
  1 sacra
• The pressure is sector wide, not company specific. Moonshot is described as being pulled into the same price war after DeepSeek pushed pricing to 0.1 yuan per million tokens. ByteDance adds another layer of pressure because it can price AI close to cost and make money elsewhere in its ecosystem.
  2 sacra 1 sacra
• Open weights are the other escape valve. Once a strong model is downloadable, developers can run it through clouds and inference platforms instead of paying a premium proprietary API. That shifts buyer behavior from picking a single model vendor to shopping for the cheapest reliable hosting layer.
  3 axios 4 sacra

The next phase is likely to split the market in two. Commodity text and reasoning will keep getting cheaper, while value moves toward distribution, multimodal workflows, and apps that own demand. For MiniMax, that makes Talkie, Hailuo, and bundled audio, video, and agent products more strategically important than winning a pure token pricing battle.