September 22, 2026
The open-model leaderboard has changed: Xiaomi’s MiMo-V2.6-Pro scored 46 on the Artificial Analysis Index
The index jumped from 26 to 46 in one generation, while the API price stayed the same: $0.435 per million input tokens
Artificial Analysis
@artificialanlys
MiMo-V2.6-Pro debuts as the top open-weight model in the Artificial Analysis Intelligence Index (46). At $0.13 per index task, it reaches the Pareto frontier of “intelligence versus cost per task” @Xiaomi has just released MiMo-V2.6-Pro, an open-weight model that makes a major leap in intelligence over its predecessor MiMo-V2.5-Pro (Intelligence Index: 26). Its pricing remains just as appealing: $0.435 per 1M input tokens (with a 99% discount on cache hits) and $0.87 per 1M output tokens. That makes MiMo-V2.6-Pro one of the most cost-effective models to deploy. MiMo-V2.6-Pro is built as an MoE: 1.02T total parameters, 42B active. We will publish a further analysis of the model soon. See the full MiMo-V2.6-Pro benchmark analysis here: https://artificialanalysis.ai
· 858.9K views
On the code-agent benchmark DeepSWE v1.1, MiMo-V2.6-Pro scored 71.9 versus 19.0 for the previous version. The price did not change.
The open weights have come very close to closed flagships. On Terminal Bench 2.1, Xiaomi measured 89.9 for itself, 89.1 for Claude Opus 5, and 88.8 for GPT-5.6 Sol.
The price did not move. $0.435 per million input tokens and $0.87 per million output tokens: the same as its predecessor with an index of 26. Cached reads cost $0.0036 per million, a 99% discount.
It is not a clean sweep. On Terminal Bench 4.0, MiMo scores 34.9 versus 49.0 for Opus 5; on ExploitBench, 47.9 versus 70.0.
**Its own hardware cannot handle everything.** The weights are on HuggingFace under the MIT license, commercial use is allowed, but the vendor recommends deploying the flagship on two nodes with 8 GPUs each: it does not fit on a single eight-GPU machine. For ordinary hardware, the same release includes MiMo-V2.6-Distill-Qwen-9B, where SWE Pro rose to 44.6 from 32.0 for the base Qwen3.5-9B, and it launches with one line.
Without your own hardware, OpenRouter already serves the model under the ID xiaomi/mimo-v2.6-pro: a 1 million-token context window and up to 131 thousand tokens in the response.
On the same date, Xiaomi released two more models in the series: flash at $0.14 and $0.28 per million tokens, and ultraspeed at $4.35 and $8.70. The vendor priced its promise of “up to 20 times faster” at exactly a tenfold markup.
Artificial Analysis promises a separate analysis of MiMo-V2.6-Pro.
