MiMo V2.6 Pro: API Pricing, Benchmarks and How to Use It

MiMo V2.6 Pro is live. Xiaomi's September 22 announcement confirms the release of Pro and Flash, with API access and downloadable weights. If you found an older page saying the model is still in training, that advice is now out of date. Xiaomi's release announcement.

The useful question is what to put it to work on. Our starting recommendation: try Pro on jobs where a wrong answer creates more work, compare Flash on routine tasks, and price UltraSpeed against time actually saved. A cheap token rate is welcome; a finished job that needs little cleanup is what pays off.

This is a source-based launch guide, checked on September 22, 2026. Mengbi has not run paid API benchmarks for this article. The cost examples below are calculations, and the suggested tests are a plan you can use with your own work.

MiMo V2.6 Pro at a glance

Sources: official model specifications and Xiaomi's model card. The checkpoint name is a download identifier; the shorter lowercase ID is what you send to Xiaomi's hosted API.

The distinction between input and output matters. Understanding a video does not mean this endpoint directly returns a generated video file. For a production workflow, specify the artifact you need: a timestamped summary, a script, a patch, or instructions for another tool. That gives you something concrete to verify.

Likewise, a large context window gives you room, not a reason to paste everything. Start with the files and evidence relevant to the job. Keep a record of which inputs produced the answer so a reviewer can follow the result without reopening an entire repository.

API pricing: Pro, Flash and UltraSpeed

The following are Xiaomi's pay-as-you-go rates per one million tokens, checked on September 22. CNY and USD are separate published regional price lists, not currency conversions performed by Mengbi. Official pricing.

Pro and Flash also support batch inference at half their corresponding token rates. UltraSpeed does not support batch. Web search is billed separately, and a Token Plan subscription does not share its quota with pay-as-you-go balance. Check the Token Plan terms before choosing a key and endpoint.

A budget you can actually check

Suppose one Pro job uses 100,000 uncached input tokens and 10,000 billable output tokens:

0.1 × ¥3 + 0.01 × ¥6 = ¥0.36

One thousand identical jobs would cost ¥360 in model tokens. This example excludes search, retries and other services. If the job needs more reasoning output than you budgeted, use the larger billed token count, not just the length of the final answer.

Now assume 80,000 input tokens are billed as cache hits and only 20,000 are uncached:

0.08 × ¥0.025 + 0.02 × ¥3 + 0.01 × ¥6 = ¥0.122

That is an illustration, not a promised cache-hit rate. Check actual usage records. A repeated-looking prompt is not enough evidence that the provider charged it as cached input.

For an agent that tries several times, add every attempt. Divide total spending by accepted results. This keeps a low sticker price from hiding the cost of failed runs and manual repairs.

Which version should you choose?

Start with Pro for difficult, multi-step work. A useful trial is an existing bug with a known reproduction, or a research question with source documents and a checkable answer. Define completion before running it: passing tests, cited passages, or an output file a colleague can open.

Give Flash the routine workload first. Document tagging, short extraction and straightforward edits are good candidates for a comparison. If both models meet the same acceptance standard, keep the less expensive one. If Flash repeatedly needs a second pass, count that work before deciding it wins on price.

Buy UltraSpeed for a measured latency problem. Xiaomi advertises up to 20× speed, while the published token rates are 10× standard Pro. Neither number tells you how long your whole task will take. UltraSpeed model page.

An interactive assistant may benefit from faster output. An overnight document batch may not. Measure time to a useful answer and total task duration, including external tools. If the slowest step is a test suite or a search request, faster text generation only fixes part of the wait.

If you are comparing across providers, our Grok 4.7 API guide and DeepSeek V4.1 Flash review cover other options. Keep the same tasks and scoring rules when testing them; headline prices alone do not make a fair comparison.

Benchmarks: useful evidence, with limits

Artificial Analysis lists MiMo V2.6 Pro at 46 on its Intelligence Index in the snapshot checked for this article. That is independent evaluation evidence. It is not a success rate for your application, and the live index can change.

Xiaomi also publishes the following results in its model card. These are vendor-reported results, not tests conducted by Mengbi.

Read across each row, not down the column. Different benchmarks test different jobs; their scores are not interchangeable. The table supports a reason to evaluate the new models on software and tool-driven tasks. It cannot tell you whether a particular change will pass your tests on the first try.

For a coding assistant, record how often it fixes the problem without creating another one. For document work, sample every numerical claim and source link. A polished answer with one invented citation can take longer to repair than a rough answer with correct evidence.

The launch also emphasizes reinforcement learning: practicing tasks, receiving feedback and updating the model during training. Treat that as a description of the development process. It does not establish that the hosted model rewrites its own weights while handling your chat.

How to call the MiMo V2.6 Pro API

1. Create a key in Xiaomi's platform and choose the pay-as-you-go service for this example.
2. Install the Python client with pip install openai.
3. Set MIMO_API_KEY in your environment. Keep the key on the server, outside browser code and version control.
4. Run a small request before connecting the model to an agent.

import osfrom openai import OpenAI
client = OpenAI(    api_key=os.environ["MIMO_API_KEY"],    base_url="https://api.xiaomimimo.com/v1",)
response = client.chat.completions.create(    model="mimo-v2.6-pro",    messages=[{        "role": "user",        "content": "Give me three acceptance checks for a CSV import feature.",    }],    max_completion_tokens=1024,    extra_body={"thinking": {"type": "disabled"}},)
print(response.choices[0].message.content)print(response.usage)

This example follows Xiaomi's documented OpenAI-compatible format and disables thinking for a short connection check. It has been syntax-checked, not executed against a paid account. For reasoning tasks, set thinking.type to enabled and choose a suitable output budget. Thinking guide.

Thinking is enabled by default. Xiaomi says max_completion_tokens covers both reasoning and the final answer, so a small limit can leave too little room for the answer. In thinking mode, the service also overrides temperature and top_p with its recommended values of 1.0 and 0.95. Check these details when migrating a client with custom sampling settings.

When adding tools, test the entire round trip. The model requests a function, your application executes it, and your application returns the result. Xiaomi's Chat Completions reference requires preserving reasoning_content in the history of multi-turn thinking tool calls. A wrapper that drops unfamiliar fields can break this exchange.

Protocol compatibility gets the first request through. It does not validate your streaming parser, JSON checks, tool permissions or retry policy. Check those in the client you will actually ship. For choosing that client, see our AI coding agents comparison.

Open weights and upgrading from V2.5

The MIT-licensed Pro-RL checkpoint is available from XiaomiMiMo on Hugging Face. Its 42B active parameters describe computation per token; the model card still lists 1.02T total parameters. Do not size a deployment as if you were loading a standalone 42B model. Follow the card's deployment guidance and measure memory use with your intended context and concurrency.

For teams without a model-serving operation, an API trial is the more direct way to assess usefulness. Self-hosting is a separate infrastructure decision involving model storage, accelerator capacity, serving software and ongoing operations. Downloadable weights make it possible to investigate that option; they do not make it free to run.

There is also a concrete migration deadline: Xiaomi's pricing page says mimo-v2.5-pro and mimo-v2.5 retire on October 21, 2026, at 10:00 Beijing time. Plan the move before that date. Retirement notice.

Test mimo-v2.6-pro as the replacement for V2.5 Pro, and evaluate mimo-v2.6-flash for jobs currently using mimo-v2.5. Keep the existing prompts initially so you can identify changes caused by the model. Then adjust prompts where the evidence shows a need.

Roll out to a small share of work first. Check accepted results, spend, response time and failures. Before the retirement date, choose a fallback that will still be available afterward; an old model scheduled for shutdown is not a lasting recovery plan.

A practical acceptance test before rollout

Build a small test set from work your team recognizes. Include easy tasks, difficult tasks and at least a few inputs that previously caused trouble. This is our suggested method, not a benchmark result:

Decide the acceptance rules before seeing which model produced which answer. Keep rejected runs in the cost calculation. If a team member needs to rewrite the output, record that too.

Our recommendation is to start with one troublesome recurring job. Give Pro and Flash the same inputs, inspect their work, then expand the winner's role. That will tell you more about fit than another model ranking. For a broader reference point on complex work, read our GPT-6 Astra API guide.

Frequently Asked Questions

Is MiMo V2.6 Pro available now?

Yes. Xiaomi's September 22, 2026 announcement confirms availability. The official API ID is mimo-v2.6-pro; the open-weight checkpoint uses the separate name MiMo-V2.6-Pro-RL.

How much does MiMo V2.6 Pro cost?

Standard official USD rates are $0.435 per million uncached input tokens and $0.87 per million output tokens. Cached input is $0.0036 per million. Tools, retries and service choices can change the total bill.

Is MiMo V2.6 Pro free or open source?

Xiaomi publishes an MIT-licensed open-weight checkpoint. Hosted API usage has published charges, and running the weights yourself still requires infrastructure. Do not assume a free trial applies to every account or service.

Can it understand images, video and audio?

Yes. Xiaomi's Chinese model page and model card list all three alongside text input. The documented model output is text; generating other media requires a suitable workflow and tools.

Should I use Pro, Flash or UltraSpeed?

Try Pro for hard jobs and Flash for routine volume. Consider UltraSpeed after measuring a response-time problem. Its higher token price needs to earn its place in your workload.

When do the older V2.5 models shut down?

Xiaomi lists October 21, 2026, at 10:00 Beijing time for mimo-v2.5-pro and mimo-v2.5. Test the replacement and a working fallback before then.

Sources and verification

Official prices and API details are linked alongside the relevant claims. The benchmark section identifies independent and vendor-reported results; the budget examples and rollout recommendations are Mengbi's analysis.

Last updated: September 22, 2026. Sources were checked on that date. Prices, availability and live benchmark scores may change.