Xiaomi's MiMo-V2.6-Pro tops the open-weight index at $0.13 a task; Claude Opus 5 still leads on coding (2026)

Xiaomi released and open-sourced the MiMo-V2.6 series on September 21, with MIT-licensed weights for a 1.02-trillion-parameter flagship and a smaller model built for volume. 1 Both models take text, images, video and audio, return text, and carry a 1M-token context window. 2 Both are live on Xiaomi's API and OpenRouter, and both can be downloaded. 2

Signal	Confirmed detail	What it means for you
What shipped	`mimo-v2.6-pro` and `mimo-v2.6-flash`, MIT weights, a distilled Qwen-3.5-9B, the technical report, 7,000-plus RL task environments, an RL framework and mini-harnesses. 1	Two model slugs to try, or a full training stack to read.
What is new	One mixed reinforcement-learning run covering coding, general-agent, visual and cybersecurity tasks. 2	The method is the story; the architecture is a familiar sparse mixture of experts.
What it costs	Pro $0.435 and $0.87 per million uncached input and output tokens; Flash $0.14 and $0.28. 3	Flash costs a third of Pro for a few points less.
Where it lands	46 on Artificial Analysis's Intelligence Index, the highest open-weight score, at $0.13 per index task. 4	A downloadable model now matches the proprietary release of the same day.
The catch	On Xiaomi's own table Pro still trails Claude Opus 5 on the coding rows, and every agentic evaluation here was run by the vendor. 2	One band of work is still closed-model territory.

What Xiaomi's own table shows

Xiaomi's model card puts Pro at 71.9 on DeepSWE v1.1 against 19.0 for MiMo-V2.5-Pro, 53.1 on AutomationBench and 89.9 on Terminal-Bench 2.1. 2 It also claims Pro performs on par with Claude Opus 5 across most agent benchmarks, and on the general-agent rows that holds: AutomationBench 53.1 against 50.3, Terminal-Bench 2.1 89.9 against 89.1, Toolathlon-Verified 76.9 against 80.6. 2 The coding rows do not: DeepSWE 71.9 against 74.0, ProgramBench 26.5 against 37.0, MiMo Code Bench 63.2 against 68.6, Terminal-Bench 4.0 34.9 against 49.0. 2 Flash gives up little for its price — 67.9 against Pro's 71.9 on DeepSWE, 52.3 against 53.1 on AutomationBench. 5

Where it stands on someone else's scale

Artificial Analysis Intelligence Index bar chart, MiMo-V2.6-Pro marked by a red arrow at 46 — Artificial Analysis's Intelligence Index, in the copy testingcatalog published with its report on the launch: the arrow marks MiMo-V2.6-Pro at 46, the highest open-weight bar, seven points below Claude Fable 5.1 and GPT-6 Astra at 53. 4 6

The bar it lands beside is a proprietary one: SpaceXAI's Grok 4.7 scores 46 as well. 7 The same measurement puts Pro at 125.2 output tokens per second against a 77.7 median for open-weight models of its size, and 2.19 seconds to first token. 4

Before you build on it

The licence is open; running it is another matter. Pro's repository holds about 570 GB of weights, and Xiaomi's recommended SGLang launch runs it at 16-way tensor parallelism across two nodes. 2 Flash, at 309 billion parameters with 15 billion active, runs at eight-way parallelism on one node. 5

On the API, UltraSpeed — the variant Xiaomi says generates up to 20 times faster — bills at ten times Pro's rates and is excluded from batch inference. 3 Batch work on the standard models costs half price, and the previous mimo-v2.5-pro and mimo-v2.5 slugs retire on October 21. 3

Xiaomi put the training method in the open with the weights: a technical report, the RL environments and the harnesses. 1 The report puts Pro's RL post-training at $2.6 million and Flash's at $0.9 million, run across thousands of GPUs with 1,568 prompts and 16 rollouts each per step, and says confirmed reward-hacking trajectories stayed under 2% for both. 8