Xiaomi MiMo-V2.5 API Price Cut Up to 99%: Pro ¥3/¥6 vs ¥7/¥21 Pricing Comparison

How much the price actually dropped

Xiaomi MiMo-V2.5 API pricing was permanently cut starting 00:00 Beijing time on May 27, 2026, by up to 99%, and the context-length tiering was removed. The official announcement frames it as "a permanent overhaul of the entire model pricing system," applied globally at once (source).

The direct effect for you: the same call volume now produces a much smaller bill, especially on cache hits and long-context workloads.

Before-and-after pricing comparison

The official pricing page lists the old V2 series and the new V2.5 series side by side, which makes a clean pricing comparison (source):

All prices are CNY per 1M tokens. The old column is the V2 series (Pro maps to mimo-v2-pro, standard to mimo-v2-omni), which is the baseline behind the "up to 99%" claim.

The cut is not even: cache hits are the big one

The "up to 99%" refers to cache-hit input, not a 90% discount across every tier. Break it into three lines:

• Cache-hit input: Pro ¥2.80 (long context) → ¥0.025, down 99%; nearly free input.
• Cache-miss input: Pro ¥7 → ¥3, down 57%; long context ¥14 → ¥3, down 79%.
• Output: Pro ¥21 → ¥6, down 71%; long context ¥42 → ¥6, down 86%.

So the real savings depend on your workload mix. The more repeated context you have (multi-turn chats, agents re-reading tool definitions, fixed RAG documents), the higher your cache-hit rate and the closer you get to that 99% line. Purely one-shot generation mostly captures the 71%-86% output reduction.

Pricing comparison against rivals: worth switching?

Put MiMo-V2.5 side by side with mainstream domestic and international models, and its value is strong. Using flagship cache-miss input/output as the benchmark:

This pricing comparison shows MiMo-V2.5-Pro is basically in the same tier as DeepSeek V4 Pro (with a slightly cheaper input) and matches it on cache-hit pricing, while undercutting Claude Opus 4.5 by roughly 97% on input. If long-context inference cost is your bottleneck, MiMo's 1M context window plus cheap cache hits are two concrete selling points. DeepSeek and Claude figures are list prices with slightly different cache-hit definitions; check each provider's pricing page for exact numbers.

For existing users: quota reset + 5-8x volume

If you already use a MiMo Token Plan, this was not just an API price cut — the plan also got more generous. Xiaomi optimized the Token Plan billing system so the same price now yields 5-8x the usable token volume, and all active subscribers had their Credits reset at 00:00 on May 27 under the new rules (source).

Keep the two entries separate: the open API uses sk- keys and deducts account balance, while Token Plan uses tp- keys against a Credits pool — fully independent billing systems. Use MiMo API pay-as-you-go for backend services and custom apps; use MiMo Token Plan for a fixed monthly coding subscription.

Things to watch

1. The old V2 series was retired on June 30, 2026. Legacy names like mimo-v2-pro and mimo-v2-omni auto-routed to V2.5 equivalents from June 1 and went offline on June 30; new integrations should use V2.5 directly.

2. The cut removed input-length tiering, not every add-on. Web search (¥16/1K calls domestically) and ASR (¥0.5/hour) are still billed separately from token pricing; the TTS family is temporarily free.

3. Cache hits are the main cost lever. Fixed system prompts, repeated project context, and shared multi-turn history reliably capture the ¥0.025/¥0.02 cache-hit rate, pushing input cost close to free.

4. Overseas pricing is in USD. Pro maps to $0.0036/$0.435/$0.87 and the standard model to $0.0028/$0.14/$0.28 per 1M tokens, adjusted globally at the same time.