Xiaomi's MiMo-V2.6-Pro Tops Open Model Rankings at $0.13 Per Task, But Anthropic Says It Used Claude to Get There

Xiaomi has released its MiMo-V2.6 model lineup, and the flagship MiMo-V2.6-Pro now tops rankings among openly available AI models — while undercutting rival pricing by a wide margin. The release arrives days after Anthropic publicly accused Xiaomi of improperly extracting training data from Claude.

The Numbers

According to Xiaomi, MiMo-V2.6-Pro scores 46 points on Artificial Analysis's Intelligence Index, placing it ahead of open competitors including Kimi K3 and Qwen. Pricing sits at $0.435 per million input tokens and $0.87 per million output tokens. Artificial Analysis calculates that a single test task costs approximately $0.13 — a fraction of what comparably capable models charge — putting MiMo-V2.6-Pro on the so-called Pareto frontier of intelligence versus cost.

Pro is a mixture-of-experts model with 1.02 trillion total parameters, of which only 42 billion activate per request. A smaller sibling, MiMo-V2.6-Flash, ships alongside it, along with a speed-optimized variant called Pro-UltraSpeed that Xiaomi claims delivers up to 20 times the output throughput.

Reinforcement Learning at Scale

Xiaomi attributes the performance jump to a heavily expanded reinforcement learning phase, scaled along three axes: more training data per step, more varied task environments, and more compute dedicated to grading solutions. The company says the RL run took under six days and cost roughly $2.62 million for Pro and $0.85 million for Flash.

On the DeepSWE coding benchmark, Pro's score rose from 58.4 to 72.6, while Flash climbed from 48.8 to 65.7. To maintain stability at this scale, Xiaomi says it froze the model's internal distribution mechanism and layered in protections against reward hacking — shortcuts a model can exploit to score well without genuinely solving a task.

Xiaomi is also open-sourcing its RL toolkit: a technical report, the full training framework, a smaller model for further fine-tuning, and roughly 7,000 pre-built training tasks with automatic graders spanning software development, cybersecurity, office work, and web design, plus around 1,000 tasks for music composition. Code tasks draw partly from real GitHub pull requests and user queries; cybersecurity tasks pull from OSS-Fuzz, a large repository of real software vulnerabilities.

Anthropic's Accusation

The openness push contrasts with claims Anthropic raised roughly two weeks earlier in a threat intelligence report covering Claude abuse cases between December 2025 and August 2026. The report names seven Chinese AI labs — Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, MiniMax, and SenseTime — tied to campaigns Anthropic says generated roughly 190 million exchanges aimed at extracting Claude's capabilities for training rival models, a practice Anthropic terms illegal distillation.

In a case labeled GTG-16008, Anthropic says it tracked more than 400,000 exchanges over 20 days in March and April 2026, during which Xiaomi allegedly routed user conversations and coding sessions from its own MiMo models through tools called OpenClaw and OpenCode into Claude, extracting responses to enrich its own training data. Anthropic's report does not fully document the provenance of earlier teacher data used in Xiaomi's internal distillation pipeline. Xiaomi has not issued a public response to the specific allegations as of publication.

What This Means

The timing here is not coincidental. Anthropic's accusation and Xiaomi's benchmark-topping release both concern the same underlying question: where did MiMo's capability gains actually come from? Xiaomi's technical disclosures — training costs, RL methodology, opened task datasets — are unusually transparent for a frontier lab, which makes the contrast with Anthropic's distillation claims sharper. If Anthropic's case holds up, it suggests some portion of MiMo's reported gains may trace back to Claude outputs rather than the RL scaling Xiaomi is publicizing, complicating how outside observers should weigh the benchmark numbers. Either way, the price-performance ratio Xiaomi claims — a fraction of a cent per task at flagship-level scores — will pressure other open-model providers regardless of how the data-sourcing dispute resolves.