Xiaomi's MiMo-V2.6-Pro now sits at the top of the open-weight rankings and charges $0.435 per million input tokens, less than a quarter of what Anthropic and OpenAI charge for models in the same performance class.
Xiaomi released MiMo-V2.6-Pro on September 21, and the number that matters isn't the benchmark score. It's $0.435. That's what Xiaomi charges per million input tokens through its API, against $4 for Claude Opus 5.5 and $5 for GPT-5.6 Sol. Output tokens run $0.87 a million on MiMo-V2.6-Pro, compared with $20 for Opus 5.5 and $30 for GPT-5.6 Sol. The model isn't a budget also-ran, either: it scored 46 on Artificial Analysis' Intelligence Index, tying xAI's Grok 4.7 and beating Grok 4.6's 44 and Google's Gemini 3.8 Flash at 41.
We covered MiMo-V2.6's base release last week, including Anthropic's claims that Xiaomi had distilled its models off Claude's outputs. The Pro variant is a separate, larger release, and it changes the argument. This is the version Xiaomi is putting up against the frontier labs directly, and it's shipping with full weights, code, and a technical report under an MIT license, not a paper promising a future launch.
The RL run for MiMo-V2.6-Pro cost about $2.62 million, according to a report from FourWeekMBA that reviewed Xiaomi's technical disclosure: 30 training steps across roughly 750,000 trajectories, finished in under six days. Xiaomi also trained a smaller Flash variant, 309 billion total parameters against the Pro's 1.02 trillion, for around $850,000 using the same pipeline. Both figures cover only the reinforcement-learning phase. Xiaomi hasn't disclosed what pretraining cost, so the true all-in number is still unknown.
That matters because a training-cost disclosure this specific is rare. Most labs, American or Chinese, release a benchmark chart and call it done. Xiaomi published a number a competitor could actually use to estimate its own margins. Fuli Luo, a former DeepSeek researcher who now runs Xiaomi's MiMo team, led the effort, and the DeepSeek pedigree shows: MiMo-V2.6-Pro beat DeepSeek's own V4.1 Pro, which scored 36 on the same index.
Xiaomi's MiMo-V2.6 Tops Open AI Charts Amid Anthropic Distillation Claims Xiaomi's new MiMo-V2.6 model just topped the global open-weights AI leaderboard, built for roughly $2.62 million. The release came less than two weeks after Anthropic accused Xiaomi of routing over 400,000 conversations through Claude to help train it, one of seven Chinese labs named in Anthropic's latest threat report. - Xiaomi MiMo-V2.6 open weights model benchmark - Anthropic Claude distillation campaign allegations Xiaomi
The architecture explains part of the price gap. MiMo-V2.6-Pro is a mixture-of-experts model with roughly 42 billion parameters active per token out of that 1.02 trillion total, so a single query touches a small fraction of the full network. Xiaomi isn't charging a discount on a full dense model. It built something structurally cheaper to run and then priced it to match.
The margin problem this creates #
Enterprise buyers evaluating Claude, GPT, or Gemini against a model like this face a question that used to be theoretical and now isn't: does a 46 on an intelligence index bought at a fifth of the price beat a marginally higher score at five times the cost? For most production workloads, coding agents, data extraction, customer support routing, the answer increasingly tilts toward whichever model clears the bar cheaply enough to run at volume. VentureBeat called MiMo-V2.6-Pro the cheapest model Artificial Analysis tracks among the top tier, and that's the sentence that should worry OpenAI and Anthropic's finance teams more than any benchmark chart.
Here's the thing. A closed lab can defend a price premium on reliability, support, or safety tooling. It's much harder to defend a five-to-tenfold premium on raw intelligence-per-dollar when the open alternative ships its weights for anyone to inspect, fine-tune, or self-host. Sonnet 5 undercuts Opus 5.5 at $2/$10, and Gemini 3.1 Pro sits at $2/$12, so the labs already know this pressure exists. MiMo-V2.6-Pro just moved the floor a lot lower, a lot faster than anyone priced in.
None of this settles the distillation dispute Anthropic raised over the base MiMo-V2.6 release. But it does mean the conversation about Chinese open models has moved past whether they can match Western frontier labs on paper. Xiaomi's technical report and MIT license make the RL run and the resulting weights checkable by anyone with the hardware to run them. The number that will actually decide adoption isn't 46. It's $0.435, and every enterprise buyer running the math already knows it.
Also read: The 10-Year Treasury Yield Just Hit 5.12%, the Highest Since 2023 • Massachusetts Regulators Move to Examine DraftKings' Use of AI on Bettors • OpenAI Is Reportedly Building a $500 a Month ChatGPT Pro Max Tier
This article is posted in AI News, check it out for more related stories.
Anthropic Says Six Other Chinese AI Labs Did What Xiaomi Did to Claude Anthropic's full threat report reveals Xiaomi wasn't acting alone: six other Chinese AI labs, including Alibaba, DeepSeek and Moonshot, ran similar campaigns to extract Claude's outputs, with Moonshot caught secretly routing live user queries through Claude without disclosure. - chinese AI labs copying Claude without attribution - how Anthropic caught multiple companies stealing Claude
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.