Xiaomi's affordable flagship AI leads the open models, and Anthropic says Claude helped get it there

What happened

Xiaomi has released MiMo-V2.6-Pro, a new open-weight AI model that its maker positions as leading the field of openly available models while pricing it well below rival flagship systems. According to The Decoder, the model's performance was achieved through large-scale reinforcement learning training that cost approximately $2.62 million — a fraction of the budgets typically associated with frontier model development.

The launch has been complicated by allegations from Anthropic, which claims that data generated by its Claude models was used, without authorisation, in the training of MiMo-V2.6-Pro. The claim raises questions about how Xiaomi sourced some of the training material behind its cost-efficient breakthrough.

Why it matters

The core story here is about what open, low-cost AI development now makes possible. If a flagship-grade open model can be trained for a few million dollars rather than the tens or hundreds of millions typically cited by leading labs, it lowers the barrier for enterprises, governments and smaller developers to build on or fine-tune near-frontier capability without depending on a handful of expensive proprietary providers.

At the same time, the Anthropic dispute is a reminder that the economics of "cheap frontier AI" are not fully settled. Questions about how training data is sourced — and whether outputs from one company's model were used to accelerate another's — will shape how enterprises evaluate the provenance and trustworthiness of the AI systems they choose to build on, particularly as open models proliferate.

By the numbers

• $2.62 million — the reported cost of the reinforcement learning training behind MiMo-V2.6-Pro, according to The Decoder.

The Renascence take

The headline number — a flagship-grade model for a couple of million dollars — is the part everyone will repeat. The more interesting story for operators is what it signals about the coming commoditisation of AI capability, and the governance questions that commoditisation drags along with it.

Cheap, high-performing open models will keep arriving, and every one of them will invite the same due-diligence question: where did the training data actually come from? Organisations racing to adopt the lowest-cost model on the leaderboard should treat data provenance as a procurement criterion, not an afterthought — because the reputational and compliance cost of building customer-facing services on disputed training data will always outweigh whatever was saved on the licence fee.