# Xiaomi MiMo V2.6 Pro Becomes Top Open Model On Artificial Analysis Intelligence Index > Source: > Published: 2026-09-21 21:21:58+00:00 Chinese companies are continuing to trade places with each other at the top of the open model pile. Xiaomi’s MiMo-V2.6-Pro has debuted at the top of the open weights leaderboard on the Artificial Analysis Intelligence Index, scoring 46 — a staggering 20-point jump over its predecessor, MiMo-V2.5-Pro, which sat at 26. The score places the Xiaomi model sixth overall, ahead of every other openly available model and within touching distance of the closed frontier. To appreciate the scale of the leap, consider the leaderboard MiMo-V2.6-Pro is crashing. The index — which [Artificial Analysis has been revising rapidly of late](https://officechai.com/ai/artificial-analysis-updates-intelligence-index-twice-in-2-days-fable-5-1-gpt-6-astra-now-tied-for-first-place/) — currently has Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra tied at the top with 53, followed by Claude Opus 5 (51), Meta’s Muse Spark 1.3 (48), and GPT-5.6 Sol (47). MiMo-V2.6-Pro’s 46 puts it just one point behind Sol — and makes it the highest-scoring open weights model in the world. Before this release, the open-weights crown was shared by Z.AI’s GLM-5.3 and Moonshot’s Kimi K3 at 44, with [Chinese labs trading the title back and forth for months](https://officechai.com/ai/chinese-startup-z-ais-glm-4-7-model-goes-past-kimi-k2-thinking-to-become-top-open-model-in-the-world/). Xiaomi released MiMo-V2.6 in two variants — Pro and Flash — both omnimodal models advanced through scaled reinforcement learning. Alongside the open weights, Xiaomi has published the technical report, its RL environments, and training code. The company says the Pro model performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks, with stronger coding, computer use, 3D reasoning, and creative capabilities than the previous generation. Architecturally, MiMo-V2.6-Pro is a mixture-of-experts model with 1.02 trillion total parameters and 42 billion active parameters. **Xiaomi Mimo V2.6 Pro Benchmarks** The headline number comes from Artificial Analysis Intelligence Index v4.3, which incorporates ten evaluations: AA-Briefcase, GDPval-AA v2, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. But Xiaomi’s own technical report gives a more granular — and more honest — picture of where the model wins and where it still trails the closed frontier. *Coding.* On DeepSWE v1.1, MiMo-V2.6-Pro scores 71.9 — close behind Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0), and ahead of Fable 5 (70.0). On ProgramBench, however, it scores just 26.5 against Opus 5’s 37.0, suggesting that while agentic coding is strong, competitive programming remains a gap. *General agent tasks.* This is where the model genuinely punches at the frontier. It beats all three closed rivals on AutomationBench v1.0.6 (53.1 vs 50.3, 45.8, and 46.2) and edges them on Terminal Bench 2.1 (89.9, the best in the table). It ties Claude Opus 5 on Agents’ Last Exam (31.6 each), and posts a competitive 1673 on GDPval-AA 2.1 against Opus 5’s 1708. The clear weak spot is Terminal Bench 4.0, where it scores 34.9 against Opus 5’s 49.0 — a reminder that the hardest terminal-based agentic work still belongs to the closed labs. On OSWorld-Verified (82.0) and Toolathlon-Verified (76.9), it lands within a few points of the leaders without overtaking them. *Cybersecurity.* The results are mixed. MiMo-V2.6-Pro posts 94.0 on CyberGym and 80.2 on Xiaomi’s own MiMo Cyber Bench, but on ExploitBench it scores 47.9 against GPT-5.6 Sol’s 78.5, and 66.3 on SEC Bench Pro against Sol’s 79.1. Offensive security is not yet a strong suit. *Visual agent work.* On MiMo Visual Coding, the model scores 72.3 — ahead of Claude Opus 5 (70.0) and Fable 5 (69.1), just behind GPT-5.6 Sol (73.4). The throughline: MiMo-V2.6-Pro is a frontier-class *agentic* model — automation, tool use, terminal work, and computer use — that trades wins with models costing many times more, while still conceding ground on the hardest reasoning and security tasks. **Xiaomi Mimo 2.6 Cost And Pareto Optimality** The performance story is only half of it. At $0.13 per Intelligence Index task, MiMo-V2.6-Pro lands squarely on the Intelligence-vs-Cost Pareto frontier — meaning no other model offers more intelligence per dollar, and no cheaper model offers more intelligence, full stop. It is, per Artificial Analysis, one of the most cost-efficient models to deploy anywhere. The API pricing tells you why. Xiaomi has held prices flat from the V2.5 generation: $0.435 per million input tokens — with a 99% cache-hit discount — and $0.87 per million output tokens, for both Pro and Flash. That keeps MiMo-V2.5-Pro among [the cheapest capable models on the market](https://officechai.com/ai/cheapest-ai-models/), and the V2.6 generation delivers dramatically more intelligence at the identical price point. Xiaomi claims that at comparable intelligence, the Pro model costs roughly one-twentieth to one-sixtieth as much as leading international models. Against Claude Fable 5.1 — which sits at the top of the frontier but costs several dollars per task — the differential is stark: near-frontier intelligence at about a fiftieth of the cost per task. This is the “Pareto frontier push” playbook that has defined the Chinese open-model wave. Rather than chasing the absolute ceiling, Xiaomi is moving the efficiency frontier outward — and doing so with open weights, which lets enterprises self-host, fine-tune, and deploy commercially without licensing friction. The [open-weights gap with closed models has been narrowing for months](https://officechai.com/ai/kimi-k3-2-8-trillion-parameters-pricing-context-window/), and a 46 — one point off GPT-5.6 Sol — suggests that gap is now measured in weeks, not quarters. **The Bigger Picture** MiMo-V2.6-Pro’s arrival is another data point in a structural shift in who builds the world’s AI. Chinese labs — Xiaomi, DeepSeek, Z.AI, Moonshot, MiniMax, Alibaba — now dominate the open weights tier, and [the share of tokens processed by US models on neutral routing platforms has collapsed from roughly 70% to 30% over the past year](https://officechai.com/ai/share-of-us-models-being-used-on-openrouter-has-collapsed-from-70-to-30-over-the-past-year/), with [Xiaomi among the top-10 most-used providers by volume](https://officechai.com/ai/these-are-the-most-popular-ai-model-companies-on-openrouter-june-2026/). Xiaomi has also proven unusually adept at the stealth-launch playbook — its [MiMo-V2-Pro appeared anonymously on OpenRouter as “Hunter Alpha”](https://officechai.com/ai/stealth-model-ox-alpha-available-for-free-for-a-week-on-openrouter-and-opencode/) before being unmasked at launch. For enterprises, the calculus is shifting fast. A model that ties or beats GPT-5.6 Sol on most agent benchmarks, at open weights and a twentieth of the cost, makes the “economic workhorse” argument for open models nearly unanswerable for high-volume agentic workloads. The question is no longer whether open models can approach the frontier — it’s how long the closed labs can keep charging frontier prices for a lead this thin.