Xiaomi just dropped a new open source beast: MiMo V2.6 Pro SPOILER: Kimi K3, Qwen3.8-Max and GLM-5.3 trail it on the AA Intelligence Index and Xiaomi kept the API pricing unchanged here’s how↓

1/ MiMo V2.6 Pro is fully multimodal, but I focused this test on coding. I used Claude Code with the MiMo V2.6 Pro API and asked for a classic arcade space shooter that feels properly finished: 3 levels, power-ups, a boss fight, CRT glow and chiptune music. all in one prompt. the detail that stood out: the folder has zero audio files. every sound effect and the whole soundtrack are generated in code. watch the boss fight in the demo, from the WARNING alarm to the chain explosion.

2/ then I checked what it did after writing the code. it tested its own game. it screenshotted every screen (menu, boss, warning alert, victory), read them back ("title screen looks strong", "gameplay reads well"), fixed its own capture tool when the boss shot hung, then redrew the boss into a bigger dreadnought before handing it over. 16 files, 5,468 lines of code, and its own line count matched mine exactly. the full run took a little over an hour (the first playable build landed in about 20 min, the rest was self-testing plus some connection retries on my end), cost $0.32 and needed 0 follow-up prompts. it made the boss too good. the laser phase wiped me every single run, so I spent one extra prompt on a god-mode key just to film the ending 😅 that's the part I'd tune, a softer last stage.

3/ the benchmark context: MiMo V2.5 Pro: 26 MiMo V2.6 Pro: 46 that puts V2.6 Pro #1 among open-weight models on the Artificial Analysis Intelligence Index, one point behind GPT-5.6 Sol. and the cost difference is substantial: MiMo V2.6 Pro: $0.13/task GPT-5.6 Sol (max): $1.99/task Kimi K3 (max): $2.00/task GLM-5.3 (max): $2.01/task these are Artificial Analysis weighted-average benchmark costs. your own bill depends on what you run.

4/ API pricing stayed unchanged from V2.5: $0.435 per 1M uncached input tokens $0.87 per 1M output tokens that’s the part that makes the $0.32 build above especially interesting. you can actually afford to experiment and iterate

5/ Xiaomi also shared a peek at its MiMo V3 research: HySparse2 the research targets long-context agent workloads, aiming to reduce compute and memory costs while improving retrieval as context gets much larger.

my take after this build: V2.6 Pro handled a 16-file game end to end and checked its own work along the way, the boss difficulty is the only thing I'd send back. for quick game prototypes and small web apps, I'd use it again, at $0.32 a build I can afford to iterate a lot. what would you build with it first? try MiMo here:

Self-testing via screenshots and fixing its own code for 32 cents is absurdly good for an open weight model.

$0.32 for a full build and iteration is kinda crazy, This moves fast.

@nrqa__ interesting drop! one thing I'd keep an eye on is how their real-world performance stacks up against the index scores. sometimes there's a gap.

MiMo V2.6 Pro feels built for doing, not just answering.

MiMo handled a lot in one go

Live on as well. Insane that they kept pricing flat

this space moves sooooo fast

you do know you could do this, like two years ago, at this quality?

keeping the API pricing unchanged while staying competitive is a smart move

damn xiaomi cooked hard with this one.. qwen has been real quiet since this dropped lmao

this is new for me.

Many people still understimate this ai model from Xiamo MiMo V2.6 Pro great model for work tbh

this is called gold

Xiaomi is pushing open models forward fast.

The self-testing and fixing loop is more interesting than the benchmark score. That changes how coding agents feel.

kept api pricing the same while beating kimi k3 and qwen3.8-max? wild flex by xiaomi.

I like the way you explain