In their published benchmarks, the model not only beats GPT-5.5 and Opus 4.8 in most tests, but sometimes goes right up to Fable and GPT-5.6 capabilities — at times even beating them outright, as on Arena.ai’s Code Arena.
It also drops eyebrow-raising scores in self-published benchmarks in coding, general agent use, knowledge work and visual tasks, mostly just right below the top tiers.
It does this while being a 2.8 trillion parameter open weights model, that will be downloadable — and tinkerable — on July 27th, they say.
Low cost per task
The cost of API use is also very low, ticking in at $3 per million input tokens and $15 per million outputs. That puts the cost of frontier-level reasoning at a new low, well below Fable’s $10/$50 and GPT-5.6’s $5/$30.
Some do say that it uses more tokens than the frontier models, somewhat muddying these waters, but according to Moonshot AI’s own numbers, the cost per task in select benchmarks comes in very low.
The model is available as of now, in mobile and desktop apps, API, and on the web at kimi.com.
Moonshot AI, the maker of the Kimi models, was accused of distillation from Claude in February. That means it allegedly used Claude’s outputs in its own training.
Read more: Kimi’s introduction, Axios and VentureBeat. Discussion on Hacker News and r/Singularity.