The short answer
Xiaomi MiMo V2.6 is the best intelligence-per-dollar in AI right now. The family has two main models, both with open weights under an MIT license and a 1-million-token context window:
- MiMo-V2.6-Pro — the flagship. Artificial Analysis scores it 46 on the Intelligence Index, the top open-weights score on the board (tied with Grok 4.7), at an AA-measured $0.13 per task.
- MiMo-V2.6-Flash — the volume pick. Index 38 at $0.06 per task in our October 8 ranking — the cheapest per-task figure of any model we track.
The comparison that matters, all from the same independent benchmark:
| Model | AA Intelligence Index | Cost per task | Verdict |
|---|---|---|---|
| MiMo-V2.6-Pro | 46 | $0.13 | Smarter than Haiku 5.5 and cheaper per task |
| Claude Haiku 5.5 Max | 43 | $0.21 | Outscored and undercut |
| MiMo-V2.6-Flash | 38 | $0.06 | Same intelligence as Luna Max, cheaper |
| GPT-6 Luna Max | 38 | $0.07 | Matched on brains, beaten on price |
Index scores via Artificial Analysis (Pro figures from AA's published model page; Flash/Luna/Haiku from AA Intelligence Index v4.3.2, the dataset behind our 7-model ranking). We didn't measure these — Artificial Analysis did. We're just reading their leaderboard out loud.
Sources: Artificial Analysis: MiMo-V2.6-Pro — Intelligence Index 46, $0.13/task (independent benchmarks); Xiaomi MiMo Open Platform (official console + API pricing); VentureBeat: MiMo-V2.6-Pro debuts as the top open-weights model; Tech-ish: MiMo-V2.6-Pro is the top open AI model you can download; Eesel: Xiaomi MiMo V2.6 review — benchmarks and value. Primary sources: the Xiaomi MiMo Open Platform and Artificial Analysis.
What MiMo V2.6 actually is
Xiaomi — the phone and EV company — released MiMo V2.6 on September 21, 2026, and it's a weirder, more generous package than the usual frontier launch:
- Open weights, MIT license. The actual Pro and Flash checkpoints are downloadable and fine-tunable. No other model at this intelligence tier gives you that.
- 1-million-token context, natively omnimodal. Text, image, audio, and video in; built for coding agents, visual reasoning, and computer-use tasks.
- RL-trained for agency. Xiaomi ran ~750,000 reinforcement-learning trajectories across coding, cyber, and knowledge-work environments. On the held-out DeepSWE software-engineering benchmark, Pro went from 58.4 to 72.6 across the run (Xiaomi's numbers — vendor data, but the held-out design is the right kind of claim).
- A real API platform. platform.xiaomimimo.com serves Pro, Flash, and a Pro UltraSpeed tier, plus voice models (TTS, voice design, voice cloning, speech recognition) on the same OpenAI-compatible endpoint.
The training economics explain the pricing: Xiaomi says the Pro RL run cost about $2.6M and Flash about $850K — a fraction of a frontier training budget — and passed the savings straight through instead of pricing to the brand.
What $10 actually buys you
This is the part nobody else writes down, so let's do the arithmetic on Xiaomi's published API rates:
| MiMo-V2.6-Flash | MiMo-V2.6-Pro | |
|---|---|---|
| Input / 1M tokens | $0.14 | $0.435 |
| Output / 1M tokens | $0.28 | $0.87 |
| Cached input / 1M | $0.0028 | $0.0036 |
| $10 ≈ how many finished tasks* | ~166 tasks | ~77 tasks |
*Task counts from Artificial Analysis' measured cost per task ($0.06 Flash, $0.13 Pro). Token equivalents: $10 is ~71M input tokens on Flash or ~23M on Pro at list rates.
Read that again: ten dollars is a month of serious agentic coding work — hundreds of finished tasks, each one a multi-step reasoning job with tool calls. The equivalent intelligence on Haiku 5.5 would run you roughly 3.5x more per task. This is why the price-to-performance crown isn't close right now: MiMo isn't 10% cheaper, it's a different order of magnitude.
One honest footnote: these are reasoning models, and the thinking is billed as output tokens. A task that "costs" $0.06 includes all the model's internal deliberation. Budget for it — it's still the cheapest thinking on the market.
Two ways in: Token Plan vs pay-as-you-go
Xiaomi sells access two ways. Pick the one that matches how you work:
- Token Plan — $6/month for 4B tokens. A subscription aimed at coding tools: a flat monthly fee, a huge token bucket, use it at your own pace. If you live inside a coding agent, this is absurdly cheap — 4 billion tokens for the price of a coffee. Note it uses its own endpoint and key type, and it's designed around coding workflows, so check the plan terms in the console for your use case.
- Pay-as-you-go credits — our route. Load credits onto your account, generate an API key in the console, and point anything OpenAI-compatible at
https://api.xiaomimimo.com/v1. You spend exactly what you burn, per the table above, across every model on the platform — Pro, Flash, voice, all of it. No subscription, no waste.
Either way the setup is five minutes: Xiaomi account → console → (plan or credits) → key → paste into your agent, Cursor, or script. The API speaks OpenAI's dialect, so existing tooling just works.
The invite: 10% off your first Token Plan
If you go the Token Plan route, don't pay full price on day one. Xiaomi's invite program takes 10% off your first paid Token Plan order:
- Sign up at the Xiaomi MiMo Open Platform.
- Within 3 days of signing up, enter the invite code 8MK7HP in the console sidebar. Older accounts can't redeem a code.
- Buy your first Token Plan within 30 days: the 10% comes off at checkout and stacks with Xiaomi's own first-purchase discount. API top-ups don't count.
- Skip it if you already got a free month of Token Plan: Xiaomi's terms exclude those accounts from the discount.
Reader offer
Try MiMo V2.6 with 10% off your first Token Plan
Sign up on the Xiaomi MiMo Open Platform and enter the invite code 8MK7HP in the console sidebar within 3 days of creating your account. The 10% discount then applies at checkout on your first paid Token Plan, as long as you buy within 30 days.
Claim 10% off with code 8MK7HP
Disclosure: this is our referral code. If you buy a Token Plan with it, Xiaomi credits us 10% of your first payment as MiMo API credits (not cash). It never changes our verdict. Official program terms.
The honest caveats
- Pro is not the fastest. Artificial Analysis measures ~46 tokens/sec with ~4s to first token on Pro — fine for agents, not for latency-sensitive chat. Flash is the speed pick; Pro UltraSpeed exists at 10x the price for those who truly need it.
- Thinking costs output tokens. The headline rates look even better than they are until you remember the reasoning trace is billed. It's still the cheapest reasoning available — just don't compare input-only math.
- Token Plan is coding-oriented. It uses a separate endpoint and key type from pay-as-you-go. If your workload isn't agent/coding-shaped, read the plan terms before committing.
- Data goes to Xiaomi's platform. Xiaomi is a Chinese company; API traffic is served on its infrastructure. For personal projects and most commercial work that's a non-issue — for regulated or sensitive data, the open weights let you self-host instead. Know your threat model.
- Benchmarks aren't your workload. An index of 46 is a composite across AA's suite. Run 30–50 of your own real tasks on Pro before moving anything production off your current model.
What to do this week
- Load $10 in credits and run your own bake-off. That's ~166 Flash tasks or ~77 Pro tasks — enough for a real evaluation, for less than lunch.
- Default to Flash, escalate to Pro. Flash handles the volume at $0.06/task; Pro takes the tasks that actually need the extra 8 index points.
- Heavy coding-agent user? Do the Token Plan math: $6 for 4B tokens is the cheapest serious coding subscription in AI right now — and the invite code above takes another 10% off the first order.
- Track our ranking. When Artificial Analysis refreshes its index, we'll update the cost-per-task ranking — if anyone dethrones MiMo on value, you'll read it there first.
Frequently asked questions
Is Xiaomi MiMo V2.6 really better than Claude Haiku 5.5?
On price-to-performance, yes — and it's the benchmarks saying it, not us. Artificial Analysis scores MiMo-V2.6-Pro at 46 on its Intelligence Index versus 43 for Claude Haiku 5.5 Max, while measuring $0.13 per task for Pro against $0.21 for Haiku. Higher score, lower cost per task: that's a clean win on both axes.
Is MiMo V2.6 better than GPT-6 Luna?
MiMo-V2.6-Flash matches GPT-6 Luna Max exactly on the AA Intelligence Index (38 vs 38) at a lower measured cost per task ($0.06 vs $0.07). The Pro version pulls ahead decisively at index 46. So: equal-or-better intelligence for less money, whichever tier you pick.
How much does the Xiaomi MiMo API cost?
Pay-as-you-go list rates on Xiaomi's platform: Flash $0.14 per million input tokens / $0.28 output; Pro $0.435 / $0.87. Cached input is ~99% cheaper ($0.0028–$0.0036 per million). A Token Plan subscription is $6/month for 4B tokens, aimed at coding tools.
How do I get the 10% off MiMo Token Plan?
Sign up at platform.xiaomimimo.com and, within 3 days of creating the account, enter the invite code 8MK7HP in the console sidebar. Buy your first Token Plan within 30 days and the 10% comes off at checkout, stacking with Xiaomi's first-purchase discount. Accounts that already received a free month of Token Plan are excluded (official Refer & Earn terms, updated August 13, 2026).
Is MiMo V2.6 open source?
The weights are open under an MIT license — downloadable, fine-tunable, self-hostable. The hosted API is Xiaomi's commercial service. Few models at this intelligence level give you both options.
What's the catch with MiMo V2.6?
Three: Pro is slower than speed-focused rivals (~46 tok/s, ~4s to first token per AA); reasoning tokens are billed as output, so budget for the thinking; and API traffic runs on Xiaomi's infrastructure, which matters if you have data-residency constraints — self-hosting via the open weights is the alternative.
Want the short version for a specific app? Start with our full directory of 13 reviewed tools and apps or see how we rank them.