# Xiaomi MiMo API Pricing: Rates and Limits (2026) By Osei Tetteh · 2026-10-30 · Source: https://www.activepieces.com/blog/xiaomi-mimo-api-pricing-rates-and-limits-2026 --- The per-million-token list price is the least useful number on Xiaomi's pricing page. The bill's set by the plan you buy, the cache and rate-limit terms, and how much of your traffic is output, which is why teams wiring the model into a workflow tool like [Activepieces](https://www.activepieces.com) tend to see costs that look nothing like the headline figure. The MiMo API's Xiaomi's hosted access to its MiMo model family, sold per model by token or through Token Plan subscriptions. ## Xiaomi MiMo API pricing: every plan and model ### The full price table, checked 2026-10-30 As of 2026-10-30, the table covers only the models and plans whose prices could be sourced. The only prices I could source are the two output rates, taken from the vendor pricing pages compared later in this article. The "not confirmed" cells matter as much as the filled ones: each's a term you must get in writing before a budget depends on it. | Item | Price | Inclusions | Context window | Limits | |---|---|---|---|---| | MiMo-V2.6-Flash | $0.28 per M output tokens; input and cached input not confirmed | Per-token API access | not confirmed | not confirmed | | MiMo-V2.6-Pro | $0.87 per M output tokens; input and cached input not confirmed | Per-token API access | not confirmed | not confirmed | | Token Plan Lite | not confirmed | Subscription token allowance | not confirmed | not confirmed | | Token Plan Standard | not confirmed | Subscription token allowance | not confirmed | not confirmed | | Token Plan Pro | not confirmed | Subscription token allowance | not confirmed | not confirmed | Only output rates were found. Input, cache, batch and plan prices are unconfirmed, so this snapshot ranks the models on output price and nothing more. It also gives you the list of questions that decide cost, and you answer those by pricing one workload, not one model. ### Which model is cheapest for which job On output price alone, **Flash is cheapest at $0.28 against Pro's $0.87**. Without input and cache rates I can't call it cheapest per finished task, and picking on the model name alone would be a guess. The naming suggests Flash is the lighter tier and Pro the heavier one, but a lighter model that produces longer answers, or needs retries, can cost more per finished task. Start with your own job: its input size, its output length, and how often the same prompt prefix repeats. Then test both models against that shape. ## Fees and limits behind the MiMo list price Six terms move a MiMo bill more than the per-million-token rate does: - The cached-input discount - Batch pricing - Rate limits - Free credits - Billing method - Whether reasoning counts as output Carry every one the page doesn't state into procurement as a question, because it's a line you can't price yet. ### MiMo API cache and batch discounts explained In agentic coding, [cache reads are 97.05% of billed tokens](https://doi.org/10.5281/zenodo.22688704), cache writes 2.57%, output 0.38% and fresh input 0.01%. So the cached-input rate sets nearly the whole bill, and the headline input price applies to a sliver of traffic. ![A loaf almost entirely one colour with tiny crumbs of other colours](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/c47da2c1-7fe3-4aff-806d-43f13383cee5/xiaomi-mimo-api-pricing-rates-and-limits-2026-il-d11e357c.webp) ![Where billed tokens go in agentic coding sessions](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/37145712-9be6-413e-b2b6-37c9494f788e/xiaomi-mimo-api-pricing-rates-and-limits-2026-st-4659bdf0.svg "Source: Zenodo") That split comes from one agentic-coding dataset. It won't hold for a support drafter or a file summariser, where far less of the input repeats, so measure your own mix. If the page doesn't state a cache discount, you can't estimate a coding-agent workload at all. When an agent makes many small tool calls, it hits the request cap before the token cap on the Pro tier, which allows 100 requests per minute and 10M tokens per minute. ![A huge empty pipe ending in a small turnstile with a queue of cups](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/55bee10d-855b-45e6-ac87-2a8b9cf6647a/xiaomi-mimo-api-pricing-rates-and-limits-2026-il-f2ff0b96.webp) Spread evenly, that's 100,000 tokens per request. Treat free credits as evaluation budget. They let you test the workload's shape, but they say nothing about steady-state cost or whether the limits above survive production. ### How reasoning tokens are billed **Only 0.38% of billed tokens are output**, but it's usually the dearest rate per token, so a small volume of premium-priced tokens drives a disproportionate share of your bill. **If the page counts reasoning tokens as output**, a reasoning-heavy model can cost far more than its output share suggests. **Until the page says otherwise, price reasoning at the output rate, which gives you the worst case.** That worst case is easiest to judge against where the tokens actually go in a coding session, and the chart below shows it. ## Three named workloads and their monthly costs To cost a workload, multiply its monthly token volume by the rate for the model you pick. The table sets out that arithmetic for all three, with the per-request token counts left for you to measure. | Workload | Input tokens | Output tokens | Flash cost | Pro cost | Pro cached | Arithmetic | |---|---|---|---|---|---|---| | Support drafter (10,000 tickets) | 10,000 × prompt plus ticket length | 10,000 × reply length | Input and output at Flash rates | Input and output at Pro rates | Repeated prompt share at Pro's cache-hit rate | 10,000 × (in × rate + out × rate) | | File summariser (2,000 files) | 2,000 × file length | 2,000 × summary length | Input and output at Flash rates | Input and output at Pro rates | Only shared instructions at the cache-hit rate | 2,000 × (in × rate + out × rate) | | Coding assistant (5 devs) | 5 × sessions × context resent per turn | 5 × sessions × generated code and tool calls | Input and output at Flash rates | Input and output at Pro rates | Resent context at the cache-hit rate | 5 × sessions × (in × rate + out × rate) | Because the system prompt and policy text repeat on every ticket, the support drafter is the workload where caching changes the Pro line most. Replies are short, so input usually dominates, and the cache-hit rate matters more than the output rate. For the 2,000 long documents, every file's different, so little of the input can be cached beyond shared instructions. Input dominates and output's small, which makes the input rate the number to compare between Flash and Pro. For the coding assistant, agents resend context on every turn and generate code and tool calls, so both sides of the bill grow. Caching helps the resent context, but output-heavy sessions keep the output rate in play, so measure a real week of sessions before choosing a tier. ## Hidden costs and limits that surprise MiMo users Expect surprises in four places: reasoning tokens billed as output, retries after throttling, model changes, and where your data is processed. ### Reasoning and output-token inflation If reasoning is billed as output, the same task can cost very different amounts. [TaskCents](https://taskcents.com/models/mimo-v2-6-pro/) measured 1,217 to 4,535 output tokens to translate 1,000 words, so budget for the high end, not the midpoint. ![How much billed MiMo output is hidden reasoning](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/9c30afca-41a8-4e5b-b04e-606a17e93613/xiaomi-mimo-api-pricing-rates-and-limits-2026-sh-bd53cebd.svg "Source: taskcents.com") Stretch to a blog post and the range is 442 to 1,790 tokens, a fourfold spread. Product descriptions ranged from 531 to 1,080, and proofreading a page from 416 to 656. On short tasks the numbers barely move: answering a question took 451 to 482 tokens, and a customer email 187 to 262. Price long generation at the top of its range, and cap output length per request. ### Throttling and retry costs Because agentic sessions resend their whole context on every call, a retry repeats the largest part of the bill. Across 722 coding sessions and 150,902 calls, cache reads made up 97.05% of billed tokens, as the chart shows. Retries therefore mostly re-bill cached context, which is cheap per token but huge in volume. Use exponential backoff with jitter, and confirm in your own usage logs whether a call that fails midway is still billed. When a model changes under a name you already depend on, your cost per task moves with it. Pin the exact model identifier, read the changelog on a schedule, and rerun your sample week before accepting any new default. Check where requests are processed and how prompts are retained before sending customer data. If the terms don't fit your obligations, route only non-sensitive workloads to MiMo. ## MiMo against DeepSeek, Qwen and GPT-6 Luna Route non-sensitive traffic to MiMo and the saving is largest on the Flash tier. At $0.28 per million output tokens, V2.6 Flash is the cheapest of the models compared here, while V2.6 Pro is not. Look at output price first, since it separates these models most. The table shows it as listed on [OpenAI's GPT-5 mini page](https://developers.openai.com/api/docs/models/gpt-5-mini) and the other vendors' pricing pages. Read it as a ranking, not a quote. | Model | Output $ per M tokens | |---|---| | MiMo V2.6 Flash | 0.28 | | Qwen3.8-Flash | 0.42 | | MiMo V2.6 Pro | 0.87 | | GPT-5 mini | 2.00 | Before trusting any row, check cache discounts, rate limits and free-tier terms for the plan you'll actually buy, because they differ by plan. ![Six models listed with output price bars rising from shortest to longest](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/0bbed201-df9b-475c-85a8-f1ed14a979fb/xiaomi-mimo-api-pricing-rates-and-limits-2026-il-6a1b0641.webp) Take a support drafter that writes 10 million output tokens a month. On output alone, Flash costs $2.80, Qwen3.8-Flash $4.20 and Pro $8.70. ### When MiMo is not the cheapest choice Pro is the case to watch. **At $0.87 it costs roughly three times Flash** and about twice Qwen3.8-Flash's $0.42. Choosing Pro for quality therefore means paying more than a rival small model. If drafts of Flash quality pass your review, Pro's money spent on nothing. If they don't, price Pro against Qwen3.8-Flash's $0.42 and test both on your own tickets. ## Calling the MiMo API from Activepieces workflows Whether a job runs on Flash at $0.28 or Pro at $0.87 per million output tokens is the one pricing decision left to you, and a platform that resells you a model has usually settled it before you open the product. ![Dot-matrix style chart visualizing OpenAI API model pricing costs for 2026](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/0c85109d-ca60-481a-becd-187511c96e7c/exa-3-0-search-api-pricing-and-rate-limits-2026-33f3bee7.webp) [Activepieces](https://www.activepieces.com/pricing) leaves that choice with you: on Plus ($20 a month, bring your own AI keys) every AI step is 1 credit, and the MiMo tokens are billed to your Xiaomi account at Xiaomi's rate, with no reseller's markup between you and those two numbers. Claude, ChatGPT, Cursor or any MCP client can reach it as well, so the model strategy stays yours to set. It doesn't discount anything. A workflow canvas replaces the script on someone's laptop with something the whole team can read and audit. Every week, the flow below reads rows from a Google Sheet, loops over them, sends each item to a model through an HTTP request, and asks the AI a second time to check or refine the first answer. ![A vertical workflow automation canvas showing ten connected steps integrating Google Sheets, HTTP, AI, and Gmail](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/cf13f915-2736-4d3d-bcb8-79317e8645df/xiaomi-mimo-api-pricing-rates-and-limits-2026-sc-898584ad.webp) It then writes a row back to the sheet, branches on the result, and emails a summary through Gmail. Watch the two AI calls, because every loop iteration pays for both. Read the canvas as a map of where tokens are spent, since each HTTP step is a line item you can see before it runs. Before you build, check Xiaomi's API documentation on whether MiMo accepts OpenAI-style requests. It decides whether you can reuse an existing request body or must write a new one. ### Routing jobs by model cost Put a conditional branch before the HTTP step. Short, well-defined jobs such as classifying a ticket or extracting a field go to the cheaper MiMo model. Jobs that need long reasoning go to Pro. Make the branch condition a column in the sheet, such as a job type, so the routing rule lives in data the team can edit and not in code. This is a tradeoff. A cheaper model that gets a job wrong costs you a rerun, so test the routing on your own examples before trusting it with volume. ### Logging token usage against a budget If Xiaomi's API returns usage the way the docs describe, each HTTP response reports how many tokens it consumed. Insert a row into a sheet after every call with the model, the input tokens and the output tokens. Add a running total, and a final branch that emails the team once the month crosses a limit you set. You will see output-heavy jobs here first, because the log separates output from input. That's the number the list price hides. ## The Monday checklist for budgeting MiMo spend Done in this order, six steps turn MiMo's list price into a budget you can defend to the people signing off on it. 1. Re-open Xiaomi's pricing page and write down the date you read it. A dated note tells you later whether a surprise came from your usage or from their change. 2. Run a 50-prompt test drawn from your real workload and pick the cheapest model that passes it. Choosing on benchmark reputation means you may pay for capability your task never uses. 3. Set a spending alert before any real traffic arrives, so an output-heavy overrun becomes a notification instead of a month-end discovery. 4. Measure actual input and output tokens for a week, including cache hits. Your own ratio of output to input is the figure that sets the bill, and no estimate made before launch will match it. 5. Compare the plan's rate limits with your peak hour, not your average. Throttling at peak means retries, queues, or a larger plan, and each of those has a cost the list price never shows. 6. Re-check the changelog monthly. A new model, a changed cache rule or a revised limit can invalidate steps two through five. ## Frequently asked questions Treat it as paid until the pricing page says otherwise for the specific model you'll call. Even where a free allowance exists, it's a trial quota with its own limits, so it tells you the integration works, not what production will cost. Run your named workload against it, then price that workload at the paid rate. Even though Xiaomi lists a Token Plan next to metered pricing, a plan existing doesn't mean it covers API calls. Confirm in writing whether plan allowances apply to programmatic traffic or only to interactive tools. If they don't, you're on metered billing whatever the plan's called, and your estimate should use per-token rates. **Assume prices can move at any time**, since Xiaomi publishes no schedule that we can point to. Date every estimate to the day you read the pricing page, and re-read it before each budget review. ### Can I use the MiMo API with OpenAI-compatible tools? Test it with your own client rather than trusting the label. Basic chat calls are the easy part. Tool calling, streaming and the usage fields are where compatible endpoints tend to differ. Check that reasoning tokens are reported, and whether they're billed as output, because that decides what an output-heavy job really costs. Also run a burst past your rate limit once, so you know whether requests queue or fail. Check the license and weights on each model's own release page, since availability can differ from model to model. Self-hosting swaps a per-token bill for GPU and operations cost. It only comes out ahead when your volume's steady enough to keep the hardware busy, which a named workload will show you quickly. ## Related reading - [SiliconFlow Pricing: Model Costs Per Token (2026)](https://www.activepieces.com/blog/siliconflow-pricing-model-costs-per-token-2026) - [MiniMax API Pricing in 2026 and What Each Model Costs](https://www.activepieces.com/blog/minimax-api-pricing-in-2026-and-what-each-model-costs) - [Xiaomi MiMo-V2.6: High-Reasoning Open Models for 2026](https://www.activepieces.com/blog/xiaomi-mimo-v2-6-high-reasoning-open-models-for-2026) ## References - [taskcents.com](https://taskcents.com/models/mimo-v2-6-pro/) - [OpenAI and others](https://developers.openai.com/api/docs/models/gpt-5-mini) - [Zenodo](https://doi.org/10.5281/zenodo.22688704)