Xiaomi's MiMo-V2.6-Pro Tops Open Model Rankings as Anthropic Alleges Claude Data Theft
Xiaomi's new flagship open model leads the field at a fraction of rivals' prices, but Anthropic says the training pipeline ran through Claude.
Xiaomi has taken the top spot among openly available AI models with MiMo-V2.6-Pro, a flagship system that undercuts competing releases on price by a wide margin. The company says the model was built on a large-scale reinforcement learning run that cost $2.62 million, a figure far below what Western labs typically spend on frontier training. The result puts a consumer electronics maker ahead of dedicated AI companies on the open leaderboards.
The training budget is the headline number. At $2.62 million, Xiaomi's reinforcement learning phase sits in a range that established labs have described as covering only small experimental runs, not flagship development. That gap suggests either a more efficient training recipe or access to compute and data that Xiaomi has not fully disclosed. Either way, the economics of producing a competitive open model have shifted, and the shift favors companies with hardware distribution and existing consumer reach.
The success comes with a serious allegation. Anthropic says Xiaomi siphoned training data from Claude, its own commercial model, and used that material in the MiMo-V2.6-Pro pipeline. If accurate, the claim would mean part of the model's performance traces back to a competitor's output rather than Xiaomi's own data collection. Anthropic has not published the full evidence behind the accusation, and Xiaomi has not issued a detailed response.
The dispute matters beyond the two companies. Open model releases are already judged on benchmarks that assume clean training data, and an unresolved provenance question complicates how developers should weigh MiMo-V2.6-Pro against alternatives. It also sharpens a broader fight over whether output from one model can be used to train another without permission, a question with little settled law behind it. Labs that publish open weights now face more scrutiny of where their training tokens came from.
Watch for three things: whether Anthropic files a formal complaint or releases supporting evidence, how Xiaomi answers the data-sourcing charge, and whether independent evaluators can reproduce the model's benchmark results. If MiMo-V2.6-Pro holds its position under outside testing, the price floor for open flagship models just dropped. If the data allegations stick, the model's lead may be temporary and the legal exposure considerably more durable.