What's changing: flat pricing, more tokens, no hidden fees

The key innovation is unified billing across all context lengths. Until now, the rule was that the longer the context the model processes, the more you pay — a common practice across the entire industry. Both MiMo-V2.5 and MiMo-V2.5-Pro now charge a flat rate regardless of whether you're working with a context length of 4,000 or 1,000,000 tokens.

Other changes:

• Token Plans now offer 5–8× more usable tokens for the same price.
• Simplified and more transparent billing rules — an end to complex calculations based on prompt length.
• All current users get a complete credit reset as part of a thank-you campaign.
• MiMo-V2.5-TTS remains free for a limited time — a text-to-speech model with support for emotional transitions and regional dialects.

According to the official announcement on X, the drastic price cuts are driven by continued inference optimization and more efficient server infrastructure. Xiaomi is also promising a detailed technical blog describing how it managed to reduce costs so dramatically.

How much does MiMo-V2.5 actually cost?

For context — current pricing via the OpenRouter platform, where MiMo-V2.5 has been available since April 2026:

• MiMo-V2.5 (310 billion parameters, MIT license): $0.14 per 1M input tokens and $0.28 per 1M output tokens
• MiMo-V2.5-Pro (1.02 trillion parameters, MIT license): $0.435 per 1M input tokens and $0.87 per 1M output tokens

For comparison: GPT-5.5 Pro from OpenAI costs $15 per 1M input tokens and Claude Opus 4.7 from Anthropic is around $15 per 1M input tokens. In other words — MiMo-V2.5-Pro is roughly 35× cheaper than comparable Western models on input and over 87× cheaper on output compared to Claude Opus 4.7.

Moreover, the price cuts on OpenRouter may not be final — Xiaomi announced that new pricing is still being propagated to partner platforms and final rates on individual API gateways may vary.

What MiMo-V2.5 can do and why it matters

MiMo is a family of large language models developed by Xiaomi under the leadership of Luo Fuli — a former key researcher at DeepSeek who joined Xiaomi in late 2025. The company released its first model, MiMo-7B, in April 2025, and has since expanded the portfolio to models with a trillion parameters.

MiMo-V2.5-Pro is a natively omnimodal model — capable of handling text, image, video, and audio simultaneously. With a context window of 1 million tokens, it can process extensive documents, hours-long meeting transcripts, or complete codebases in a single pass. The model is licensed under the MIT license, meaning developers can freely use, modify, and deploy it even commercially without licensing fees.

In terms of performance, according to independent benchmarks on OpenRouter, MiMo-V2.5-Pro reaches the level of Claude Opus 4.7 in agentic tasks such as ClawEval, GDPVal, or SWE-bench Pro — tests measuring the ability to independently solve complex programming and analytical tasks. The model can autonomously complete professional tasks requiring over a thousand tool calls.

China's price war escalates

Xiaomi is not the first Chinese company to dramatically slash prices on its AI models. A few days ago, DeepSeek permanently cut the price of its flagship V4-Pro model by 75 percent — to $0.55 per 1M input tokens, bringing it to roughly one-tenth the cost of GPT-5.5 Pro. MiMo is now going even further.

This trend has major implications:

• For developers and startups — access to AI models comparable to the best Western alternatives is becoming orders of magnitude cheaper.
• For European companies — models like MiMo-V2.5 with an MIT license can be deployed on their own EU infrastructure without dependence on American providers, which is crucial from the standpoint of GDPR and European digital sovereignty.
• For OpenAI and Anthropic — pressure to lower prices will continue to intensify, especially for customers who don't need premium support or enterprise SLAs.

Xiaomi is investing at least $8.7 billion in AI over three years, as CEO Lei Jun announced in March 2026. The company is positioning MiMo as a key technology for its "Human × Car × Home" ecosystem.

Availability for Czech developers and businesses

MiMo-V2.5 is available through several channels, including:

• OpenRouter — a global API gateway that provides access to the model, including payments in dollars and euros.
• Official Xiaomi API at mimo.xiaomi.com.
• Hugging Face — model weights are freely downloadable under the MIT license for self-hosting.

For Czech businesses, the key advantage is the self-hosting option — thanks to the MIT license, MiMo-V2.5 can be run on their own servers in the EU, addressing concerns about transferring sensitive data to China or the US. While the model does not directly support Czech at the level of specialized training, as an omnimodal model with an extensive multilingual corpus, it handles Czech functionally.

What's next?

Xiaomi has indicated that it will publish a detailed technical blog about the inference optimizations that enabled the massive price cuts in the coming days. This is crucial information — if Xiaomi has managed to reduce model operating costs so radically while maintaining quality, it could inspire the entire industry.

In the context of May 2026, with DeepSeek cutting prices by 75%, MiMo by 99%, and Western models remaining at high price levels, a fundamental question emerges: how much longer can OpenAI and Anthropic hold their current prices?