Xiaomi shipped the MiMo-V2.6 series on September 22 under an MIT license, and the Pro variant landed at 46 on Artificial Analysis's Intelligence Index, tying the top of the open-weights field across a class of 114 models. The framing writes itself: a phone company's research lab now sits at the frontier, and it's charging $0.13 per Intelligence Index task to get there.

MiMo-V2.6-Pro is a 1.02-trillion-parameter sparse mixture-of-experts with 42 billion active per token and a 1M-token context, priced on Xiaomi's own API at $0.435 per million uncached input tokens and $0.87 per million output. That's the headline tier. It's not the interesting one.

The interesting one is Flash: 310 billion total parameters, 15 billion active, 1M context, and $0.14 input / $0.28 output per million tokens. On the same-afternoon frontier price cuts from OpenAI and Anthropic, the closed labs came down. They didn't come down to here. Meta's Muse Spark 1.2/1.3 contributor tier sits at $0.10/$0.20, DeepSeek-V4.1-Flash off-peak at $0.15/$0.60, GPT-5.6 Luna at $0.20/$1.20, and Claude Fable 5.1 at $10/$50, roughly 180x the per-token cost of Flash for what the benchmarks suggest is a narrowing quality gap.

Flash scores 67.9 on DeepSWE v1.1 against Pro's 71.9, 52.3 versus 53.1 on AutomationBench, 61.2 versus 62.0 on JobBench, and 87.6 versus 89.9 on Terminal Bench 2.1. The two variants share the same training regime: 30 large reinforcement-learning steps, roughly 750,000 trajectories, under six days of wall-clock, and reported training costs of about $2.62M for Pro and $850,000 for Flash. Xiaomi published more than 7,000 task environments with automatic graders and reports below 2% confirmed reward-hacking trajectories in the final run.

Both models are already routable through OpenRouter's unified API at a 1.05M-token context, which is what matters for anyone actually building. The Grok 4.7 launch two days earlier kept prices flat at the top; DeepSeek's V4.1-Flash off-peak $0.15/M pricing set the previous MIT-licensed floor. MiMo-V2.6 broke it in under two weeks.

The historical parallel is the 1996 Telecommunications Act era of long-distance pricing, when incumbent per-minute rates collapsed against resellers running on the same wires. The incumbents had brand, distribution, and enterprise contracts. They didn't have a price story. Anthropic's $10/$50 tier reads, this week, like a similar bet: that the top of the market will pay for capability whatever the floor does. It's a defensible bet. It's also the bet the long-distance carriers made.

Xiaomi's MiMo-V2.6 Tops Open-Weights Index at 46 — and Prices Flash at $0.14/$0.28

Sources