Xiaomi's MiMo-V2.6-Pro leads the open-weight index at $0.87 per million output tokens

Put the two output prices side by side and the ratio is 23 to 1 ^[25]. On input it is closer to 9 to 1 ^[26]. For a team running agent loops that write long, that ratio shows up on the monthly invoice, not in the index score.

What is being sold here is an open-weight model at the top of an independent index. Xiaomi opened its announcement with "Frontier intelligence, all the modalities, built in public" ^[2]. What most teams will do is call it over OpenRouter and never touch a GPU ^[9]. TNW's account of the release did not report what hardware it takes to serve 1.02 trillion parameters, 42 billion of them active per token ^[5][34]. Capability was rarely the only reason self-hosting got ruled out.

The capability case is better than it was and still not level. Artificial Analysis has Claude Opus 5.5 at 58, twelve points above MiMo ^[18][28]. Xiaomi's own tables put Claude Opus 5 ahead of Pro on DeepSWE and on ProgramBench, and 14.1 points ahead on Terminal Bench 4.0, where GPT-6 Astra scores 59.6 ^[17][29]. Xiaomi also says Pro performs on par with Opus 5 and GPT-5.6 Sol across most agent benchmarks; that one is the company's own testing ^[19].

Flash is the tier to test first. It costs $0.14 and $0.28 per million tokens, roughly a third of Pro, and keeps the million-token context and the multimodal input ^[14]. Against Opus 5.5 on output that is a factor of 71 ^[27]. On Xiaomi's agent benchmarks it trails Pro by four points or fewer, at 67.9 against 71.9 on DeepSWE and 61.2 against 62.0 on JobBench ^[15]. On CyberGym it beats Pro, 95.1 to 94.0 ^[16].

Xiaomi reports DeepSWE twice, and the two figures do not match. The training run ends at 72.57 for Pro and 65.68 for Flash ^[13]. The benchmark table lists 71.9 and 67.9 ^[15]. That is 0.67 of movement for Pro and 2.22 for Flash ^[30]. A team choosing between the tiers on a four-point DeepSWE gap is working inside the spread of the vendor's own numbers.

The swap turns on whether the work is a long autonomous terminal session or a bounded task you can write a grader for, and on whether output tokens dominate the bill. If the work is output-heavy and bounded, Flash goes on a shadow queue behind the current model and you compare completions on real tickets. If it is long-horizon terminal work, the 14.1 points on Terminal Bench 4.0 is what the lower price costs, and that number comes from Xiaomi's own table ^[17][29]. For latency-bound work there is Pro-UltraSpeed, which Xiaomi says runs up to 20 times faster than Pro at the same quality for ten times the token price, $4.35 in and $8.70 out ^[10].

For a team that has never written an automatic grader, the reusable part of this release is the harness. Xiaomi published the technical report, the end-to-end reinforcement learning framework, more than 7,000 task environments with automatic graders, a set of composable mini-harnesses and a smaller distilled model ^[20]. The environments cover software development, cybersecurity, office work and web design, with roughly a thousand more on music composition ^[21]. Xiaomi also published the reward design, the hyperparameters, the data mixtures and the costs, saying the point is reproducibility ^[22]. One detail is worth copying into your own eval setup: early agents solved assigned bugs by fetching the fix from a later version of the package, so Xiaomi stripped build caches and future Git history from the training environments and ran a dedicated agent to hunt for remaining loopholes ^[24].