MiMo-V2.5 API Pricing Compared: Pro vs Standard — Cache-Hit Savings & Which to Choose

Two Models, Two Price Lines: The Pricing Table

MiMo-V2.5 API pricing is now simpler. After the 2026-08-06 update, text models dropped context-length tiering; Pro and Standard use flat per-unit pricing with very deep cache-hit discounts (source).

Domestic Pricing (CNY per 1M tokens)

Overseas pricing mirrors this: Pro $0.435/$0.87 (cache hit $0.0036), Standard $0.14/$0.28 (cache hit $0.0028). The web search plugin costs ¥16 per 1,000 calls domestically and $5 per 1,000 calls overseas, billed separately from token pricing.

Pro (mimo-v2.5-pro): The Heavy Lifter

1.02 trillion parameters, MoE architecture, 1M context, 42B active parameters. MiMo's flagship text model, with official claims of Claude Opus 4.6-level agent performance. From a developer's perspective, Pro is built for:

• Agent-level code generation on complex software (multi-file refactoring, project-wide requirements to implementation)
• Long-chain reasoning (10+ step multi-stage tool calls)
• Ultra-long document analysis and codebase-level understanding

The trade-off is the higher price: output at ¥6 per 1M tokens, three times Standard. If call volume is high, leverage cache hits. Pro's cache-hit input at ¥0.025 versus ¥3 uncached saves about 99%; multi-turn conversations and same-project contexts can dramatically lower cost.

Standard (mimo-v2.5): The Multimodal Sweet Spot

310B parameters, native multimodal understanding, 1M context. Weaker than Pro on pure text, but adds native image, audio, and video understanding. Input ¥1 / output ¥2 is one-third of Pro, with cache hits at ¥0.02.

Best for:

• Mixed workloads combining text, images, audio, and video (support-ticket screenshots, technical-doc illustrations)
• Day-to-day coding assistance and routine agent tasks
• Simplified architectures where one model covers multiple modalities

If your workflow includes significant multimedia content, Standard's single-model approach beats "Pro plus a third-party vision model." For complex pure-text coding, Pro performs better; but Standard is the most natural entry point for high-frequency, cost-sensitive tasks.

How to Choose: Three Typical Strategies

Strategy A: Pure Text Coding, Single Model

Day-to-day coding, code review, small agent tasks → Standard is enough (¥1/¥2). It handles most tasks at one-third of Pro's price. Switch to Pro for complex refactoring and long-chain work.

Strategy B: Light + Heavy Dual Model

High-frequency autocomplete and simple Q&A on Standard (¥1/¥2), reasoning-heavy tasks on Pro (¥3/¥6). If 80% of calls are lightweight and 20% heavy inference, the weighted average price comfortably beats all-Pro. Ideal for backend services embedding MiMo into products.

Strategy C: Unified Multimodal

If 30%+ of requests involve images, audio, or video, use Standard exclusively. Native multimodal support removes the need for separate vision/speech models. One API key, one endpoint, all media types handled uniformly.

Cost Curve: Cache-Hit Rate Determines Real Cost

Assume 1M tokens of Pro output + 1M tokens input, total cost at different cache-hit rates:

Cache hits now approach "free input." Pro input drops from ¥3 to ¥0.025 (about 99% off); Standard from ¥1 to ¥0.02 (about 98% off). The impact is limited for output-heavy coding workloads (output price stays the same), but dramatic for heavy-context-input workloads like long-document analysis and multi-file backgrounds.

Caveats

1. The V2 series was officially retired June 30, 2026. Legacy names like mimo-v2-pro, mimo-v2-omni, and mimo-v2-flash are dead; switch to the V2.5 equivalents directly.

2. Text models no longer charge 2x for 256K–1M. The 2026-08-06 update moved to flat pricing, so long-document processing no longer triggers a higher price tier.

3. Cache-hit discounts are very deep—design reusable prefixes first. Fixed system prompts, repeated same-project contexts, and shared multi-turn history all raise hit rates and push input cost near zero.

4. API pay-as-you-go and Token Plan subscription are two fully separated systems. API keys use the sk- prefix and deduct balance; Token Plan uses the tp- prefix and deducts the Credits pool. They cannot be mixed.