XIAOMI FLAGSHIPSAME PRICE AS V2.5-PROMIT WEIGHTS1.02T / 42B ACTIVE1M CONTEXT
MiMo-V2.6-Pro API Pricing
Xiaomi's new flagship kept its predecessor's price on every line: $0.435/M input on a cache miss, $0.0036/M on a hit and $0.87/M output, printed on the same table row as MiMo-V2.5-Pro. What the release adds is a Batch API at exactly half and open weights under the MIT licence: a 1.02T-parameter mixture-of-experts that activates 42B per token, reading text, images, video and audio into a 1M-token window. V2.5-Pro was retired on October 21, 2026, after a week of Xiaomi routing its requests here, so the move is a code change rather than a price change.
Input - per 1M tokens
$0.435/M
Batch $0.217/M cache miss
Output - per 1M tokens
$0.87/M
Batch $0.435/M
Cached input - per 1M tokens
$0.0036/M
Cache writes free for now 99% off
Effective - agentic blend
$0.144/M
92/8 split - 82% cache
§ 01 / TERMINAL
MiMo-V2.6-Pro API cost calculator.
Calculator pre-loaded with Xiaomi's overseas USD list for MiMo-V2.6-Pro. For work that can wait, the Batch API halves every line, and cache writes are currently free, so a high hit rate costs you nothing to build.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
MiMo-V2.6-Pro cost on real workloads.
CODING AGENT
Repo-wide bug fix
$0.018/task
LONG DOC
Reading 100-page contracts
$0.022/doc
RAG SUPPORT
Support agent ticket triage
$0.00152/ticket
RESEARCH
Research planning turn
$0.0032/turn
Estimate · xiaomi-tokenizer-estimate · ≈3.85 chars/token Auto-counts as you type
This is a chars-per-token approximation, not a real tokenizer. Actual tokens vary by language, code density, and tool-call overhead — counts are typically ±10–20% off for English prose, more for code or non-Latin scripts. For exact billing, use the vendor's official tokenizer.
Characters 522
Words 86
Tokens (estimated) 136 tokens
Cost as input · uncached $0.00006 USD
Cost as output · uncached $0.00012 USD
Cost as cached input $0.0000005 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| MiMo-V2.6-Pro Current | $0.435 cache $0.0036 | $0.87 | $0.144 agentic 92/8 | 1M | Open-weight omni-modal flagship for long agent runs |
| MiMo-V2.6-Flash | $0.14 cache $0.0028 | $0.28 | $0.0477 −3.0× | 1M | Same window and inputs at under a third of the price |
| MiMo-V2.6-Pro-UltraSpeed | $4.35 cache $0.036 | $8.70 | $1.44 +10× | 1M | The same model up to 20x faster, at 10x the rate |
| MiMo-V2.5-Pro | $0.435 cache $0.0036 | $0.87 | $0.144 retired | 1M | What this model replaces, for the record |
| DeepSeek V4 Pro | $1.32 cache $0.044 | $3.96 | $0.569 +3.9× | 1M | Frontier reasoning with a 1M context |
| GPT-6.1 Sol | $2.00 cache $0.10 | $10.00 | $1.29 +8.9× | 1.05M | Closed frontier coding and agent work |
| Gemini 3.8 Flash | $0.75 cache $0.075 | $3.75 | $0.481 +3.3× · promo | 1M | Multimodal input from a hyperscaler stack |
MiMo-V2.6-Pro has published scores on 13 benchmarks we track, and leads 3 of them. All of them are the lab's own reports; each row links to the page that published it.
| Benchmark | Score | Rank | Leader | Source |
|---|---|---|---|---|
| Toolathlon Agents and computer use | 76.9% as reported | #1 of 9 | MiMo-V2.6-Pro leads | Vendor-reported Xiaomi |
| Terminal-Bench 2 Agents and computer use · Terminal-Bench 2.1 | 89.9% as reported | #2 of 17 | DeepSeek V4.1 Flash · 90.6% | Vendor-reported Xiaomi |
| AutomationBench Agents and computer use · Main set (as reported) | 53.1% v1.0.6 | #4 of 22 | MiniMax M3.1 Flash · 62.8% | Vendor-reported Xiaomi |
| JobBench Professional work | 62% as reported | #1 of 5 | MiMo-V2.6-Pro leads | Vendor-reported Xiaomi |
| Agents' Last Exam Agents and computer use · Split not stated | 31.6% split not stated | #1 of 4 | MiMo-V2.6-Pro leads | Vendor-reported Xiaomi |
| CyberGym Cybersecurity | 94% as reported | #2 of 8 | MiMo-V2.6-Flash · 95.1% | Vendor-reported Xiaomi |
| DeepSWE Coding | 71.9% as reported | #9 of 27 | Gemini 4 Argon · 77.9% | Vendor-reported Xiaomi |
| OSWorld-Verified Agents and computer use | 82% as reported | #3 of 8 | Kimi K3 · 84.8% | Vendor-reported Xiaomi |
| ProgramBench Coding | 26.5% as reported | #3 of 7 | Kimi K3.1 Pro · 31.5% | Vendor-reported Xiaomi |
| SEC-bench Pro Cybersecurity | 66.3% as reported | #4 of 8 | Gemini 4 Carbon · 88.2% | Vendor-reported Xiaomi |
| ExploitGym Cybersecurity | 17.8% as reported | #5 of 9 | Gemini 4 Carbon · 54.8% | Vendor-reported Xiaomi |
| Terminal-Bench 4.0 Agents and computer use | 34.9% as reported | #19 of 34 | Kimi K3.1 Pro · 65.7% | Vendor-reported Xiaomi |
| ExploitBench Cybersecurity | 47.9% as reported | #5 of 7 | GPT-6 Astra · 100% | Vendor-reported Xiaomi |
MiMo-V2.6-Pro API pricing FAQ.
What MiMo-V2.6-Pro costs, what changed from V2.5-Pro, and what V2.5-Pro's retirement on October 21 means for code that still names the old id.
Q · 01 How much does MiMo-V2.6-Pro cost? +
On Xiaomi's overseas list it is $0.435/M input on a cache miss, $0.0036/M on a cache hit and $0.87/M output. The Batch API halves all three:
$0.217 / $0.0018 / $0.435. Under our 92/8 agentic blend with 82% cache hits the planning figure is $0.14/M. Xiaomi bills overseas accounts in dollars and publishes a separate yuan list for China; the two are separate price lists, not one converted into the other, so this page reads the dollar list directly.Q · 02 Is it more expensive than MiMo-V2.5-Pro? +
No. The pricing table prints
mimo-v2.6-pro and mimo-v2.5-pro on one row with one set of rates, and the launch post says the V2.6 series adopts the same API pricing as V2.5. What is new is the Batch API at half price, which V2.5-Pro never had and will not get, and published MIT-licensed weights.Q · 03 What happens to MiMo-V2.5-Pro? +
Requests that name
mimo-v2.5-pro fail since 10:00 Beijing time on October 21, 2026 (02:00 UTC), when Xiaomi retired the name. In the week before, Xiaomi routed them here automatically: from 18:00 Beijing time on October 14 they were served and billed as mimo-v2.6-pro, at the same rates, and gained this model's image, video and audio input, which V2.5-Pro never accepted. Xiaomi had first said there would be no system replacement and changed that on October 9. Code that still names the old id now needs the new name.Q · 04 What is MiMo-V2.6-Pro-UltraSpeed, and is it worth 10x? +
It is this model served for speed: Xiaomi calls it the UltraSpeed mode of V2.6-Pro, promises up to 20x the output speed with no loss of capability, and sells it under its own id at exactly 10x every rate on this page. The invite-only V2.5 UltraSpeed it replaces cost 3x its base model, so the speed premium has more than tripled. For batch work, code review or anything a person is not watching live, it buys nothing; it earns its rate only where latency is the product.
Q · 05 Are the weights open, and how big is the model? +
Yes. The weights are on Hugging Face as
XiaomiMiMo/MiMo-V2.6-Pro-RL under the MIT licence, with a technical report and the reinforcement-learning code. The model card describes a sparse mixture-of-experts with 1.02T total and 42B active parameters (384 routed experts, 8 active), a 681M-parameter vision encoder and separate audio encoders. Self-hosting sizes to the 1.02T total for memory and the 42B active for compute per token.Q · 06 What does it accept, and what are the limits? +
Text, images, video and audio in; text out. The context window is 1M tokens with up to 128K output tokens, and pay-as-you-go accounts get 100 requests per minute and 10M tokens per minute. Thinking mode, tool calls, structured output, streaming and context caching are supported. Web search is a separate charge of $5 per 1,000 calls overseas, not part of the token rate.
Q · 07 Why is the cached rate so low, and what do cache writes cost? +
A cache hit costs
$0.0036 per million, under 1% of a miss, which is why the blended rate sits so far below the input rate. Xiaomi lists cache writes as limited-time free, so building the cache costs nothing today; the vendor gives no end date and no future write price, so this page does not model one. If a write price appears, long prompts that are read only once will cost more than the figures here.