Xiaomi Ships MiMo-V2.5-Pro — A Smartphone Giant Reaches Frontier AI at $1 Per Million Tokens
Xiaomi released MiMo-V2.5-Pro on April 22, 2026 — a 1-million-token multimodal agent that resolves 57.2% of SWE-bench Pro tasks while using 40–60% fewer tokens than comparable frontier models. At ~$1/M input tokens, it's the most token-efficient frontier-tier model yet from a non-AI-native company.
Original sourceXiaomi — the Chinese consumer electronics giant better known for smartphones, smart TVs, and scooters than AI research — released MiMo-V2.5-Pro on April 22, 2026, and the result is harder to dismiss than the category might suggest.
MiMo-V2.5-Pro combines a 1-million-token context window with multimodal capability — vision, audio, and text — in a single agent-ready model. On SWE-bench Pro, it resolves 57.2% of tasks, placing it in the top tier alongside GPT-5.4 and Claude Opus 4.6. On the multimodal ClawEval benchmark, it reaches 23.8, matching Claude Sonnet 4.6 and trailing Claude Opus 4.6 by a single point.
The efficiency numbers are what make this a story beyond just another benchmark announcement. MiMo-V2.5-Pro uses approximately 42% fewer tokens than Kimi K2.6 at equivalent benchmark scores, and roughly 40–60% fewer tokens than Claude Opus 4.6 and Gemini 3.1 Pro on comparable ClawEval trajectories. That per-token efficiency maps directly to cost: the model is priced at approximately $1 per million input tokens — a significant discount to most frontier providers.
Xiaomi has been building foundation model capabilities for two years, largely out of public view. MiMo-V2.5-Pro is the clearest signal yet that consumer hardware manufacturers — with their deep cultures of supply chain efficiency, hardware-software co-optimization, and operating-margin pressure — may have structural advantages in the token efficiency competition that pure-play AI labs don't.
The obvious risks are real: Xiaomi's AI API infrastructure is essentially unproven at scale compared to AWS, Azure, or Anthropic's systems. There are no established SLAs, no enterprise compliance track record, and no community of practitioners who've pushed it in production. The API provenance will raise the same Chinese-jurisdiction flags as other models from the region. Still, the efficiency story is compelling enough that dismissing it on brand grounds alone would be a mistake for serious AI builders.
Panel Takes
The Builder
Developer Perspective
“42% fewer tokens than Kimi K2.6 at equivalent performance means real money saved in production code agent pipelines. If Xiaomi's API proves reliable under load, this should be in every agent builder's evaluation set. The 1M context and multimodal support are table stakes at this point — the efficiency is the differentiator.”
The Skeptic
Reality Check
“Great benchmarks from a smartphone company don't equal the reliability, SLAs, and compliance posture enterprise teams need. Xiaomi's AI API is essentially unproven at scale, and no amount of impressive efficiency numbers changes the fact that building production systems on an untested provider is a risk most teams can't accept yet.”
The Futurist
Big Picture
“Consumer hardware manufacturers have spent decades optimizing for efficiency under cost and power constraints that AI labs never faced. Xiaomi's token efficiency numbers may reflect that structural advantage. If this is the beginning of hardware-native AI companies competing on efficiency rather than scale, the foundation model pricing landscape is going to shift significantly.”