Chinese LLMs in November 2026: Every Flagship Priced Side by Side, Plus Two Retirements You Need to Plan Around
A complete side-by-side of the current Chinese flagship lineup, priced from the vendors' own pages: DeepSeek V4 Pro, Zhipu GLM-5.3, Alibaba Cloud Qwen3.8-Max, Moonshot Kimi K3.1 Pro, ByteDance Doubao-Seed-2.1-Pro, Xiaomi MiMo-V2.6-Pro, MiniMax-M3.1-Flash and Zhipu GLM-5.3-Flash. Includes the cheapest 1M-context tier, the full cached-input price ladder, and two deprecation signals (the Kimi K2 family and Xiaomi's MiMo-V2.5 series) that affect any roadmap built before October.
The Chinese model market just crossed a threshold that was not on anyone's calendar: every current flagship now ships a 1M-token context window, which means context length stopped being a differentiator and price became the only axis that still separates the tiers. At the same time, two product lines were retired inside a single month. This guide prices all of them from the vendors' own pages, puts the cached-input rates side by side, and flags the two migrations you should start now rather than in December.
The price ladder, normalized to one currency
Normalizing every list price to USD at roughly CNY 7.16 per dollar produces a clean ladder. The bottom rung has four models tied at CNY 0.80 input per million tokens (about USD 0.11): Alibaba Cloud Qwen3.8-Flash, Alibaba Cloud Qwen3.8-Omni-Flash, Volcengine Ark Doubao-Seed-2.1-Lite and Zhipu GLM-5.3-Flash. All four carry a 1M context window. Their output prices separate them: Doubao-Seed-2.1-Lite at CNY 2.40, Qwen3.8-Flash and Qwen3.8-Omni-Flash at CNY 2.70, GLM-5.3-Flash at CNY 2.80, and Zhipu is currently running GLM-5.3-Flash at a limited-time 50% discount that puts the real rate at CNY 0.40 input and CNY 1.40 output. One rung up: iFlytek Spark X2.6 at CNY 0.90 / CNY 3.60, the BaseYuan Lido aggregator at a flat CNY 1.00 for both directions, Xiaomi MiMo-V2.6-Flash at CNY 1.00 / CNY 2.00, and DeepSeek V4.1 Flash at USD 0.15 / USD 0.60 in its off-peak window. The flagship tier: Zhipu GLM-5.3 at CNY 8.00 / CNY 28.00, ByteDance Doubao-Seed-2.1-Pro at CNY 6.00 / CNY 30.00, Alibaba Cloud Qwen3.8-Max at CNY 12.00 / CNY 36.00, and Moonshot Kimi K3.1 Pro at CNY 13.00 / CNY 68.00 with a highspeed variant at double that. DeepSeek V4 Pro sits alongside them at USD 0.66 / USD 1.98 off-peak and USD 2.10 / USD 6.30 at peak. For a 100M-token monthly batch, the difference between the cheapest tier and the flagship tier is roughly two orders of magnitude - CNY 62.50 for DeepSeek V4.1 Flash off-peak against CNY 900 for Kimi K3.1 Pro.
The cached-input ladder decides long-context agent economics
Headline input prices get all the attention, but for agent workloads that repeatedly read the same codebase or document set, the cached-input rate is the number that actually appears on the invoice. Here the spread is far wider than the headline spread: Xiaomi MiMo-V2.6-Pro charges CNY 0.025 cached, Xiaomi MiMo-V2.6-Flash CNY 0.02, Alibaba Cloud Qwen3.8-Flash and Qwen3.8-Omni-Flash CNY 0.10, Zhipu GLM-5.3-Flash CNY 0.23 (CNY 0.115 while the 50% discount runs), ByteDance Doubao-Seed-2.1-Pro CNY 1.20, Alibaba Cloud Qwen3.8-Max CNY 1.50, and Moonshot Kimi K3.1 Pro CNY 1.30 with cache writes billed separately by TTL at CNY 13.00 for five minutes and CNY 26.00 for one hour. Note that a Kimi cache hit is charged only at the hit rate and is not re-billed for the write, so the write TTL choice matters more than the hit rate if your reuse window is short. Two Batch API discounts stack on top of everything else: Zhipu halves the standard rate for Batch calls, and Alibaba Cloud lists a Batch File rate of CNY 0.40 / CNY 1.35 for Qwen3.8-Flash and a limited-time 50% Batch Chat rate for Qwen3.8-Max.
Two product lines were retired inside October - here is what to do
Moonshot stopped opening the Kimi K2 family to new users on October 31, 2026 and plans to retire all K2 models by the end of December. That means Kimi K2.7, which still appears in many price lists, caps at 262,144 tokens against 1,048,576 on the K3 generation - a four times difference in how much context you can actually fit. Anything you built on K2 needs a migration path to K3 or K3.1 Pro now. Xiaomi retired the entire MiMo-V2.5 family on October 21 at 10:00 Beijing time, and MiMo-V2.6-Pro is the current entry point at CNY 3.00 / CNY 6.00 with CNY 0.025 cached input. Two smaller but worth-noting signals: Volcengine Ark's Agent and Coding Plan coefficient tables mark minimax-m3 as imminently retired, so confirm routability before committing a production workload to it through a third-party aggregator; and Anthropic's price table currently shows no Mythos 5.1 row, only a navigation entry for Mythos including Fable, so any Mythos pricing you have seen quoted elsewhere is a derived figure rather than a published rate.
Which one to pick, in one page
Cheapest 1M-context input with no special requirements: Qwen3.8-Flash at CNY 0.80, because it pairs the floor price with a CNY 0.10 cache and a five-million TPM ceiling. Cheapest right now including the active promotion: GLM-5.3-Flash at CNY 0.40 / CNY 1.40, understanding the discount is limited-time and the official page publishes no end date. Cheapest native multimodal: Qwen3.8-Omni-Flash at CNY 0.80 / CNY 2.70 for text, image and video, or Zhipu GLM-5.3-FlashX at CNY 2.00 / CNY 7.00 if you want a 1M window on multimodal as well. Cheapest cached input for a long-context agent loop: MiMo-V2.6-Pro at CNY 0.025, with MiMo-V2.6-Flash at CNY 0.02 if the task does not need the larger model. Cheapest batch processing: DeepSeek V4.1 Flash in its off-peak window at USD 0.15 / USD 0.60, scheduled outside the Monday-to-Friday 09:00-12:00 and 14:00-18:00 Beijing peak. Strongest domestic flagship with an open-source path: DeepSeek V4 Pro with 384K maximum output, or ByteDance Doubao-Seed-2.1-Pro at CNY 6.00 / CNY 30.00 with 1024K context. And when you need one model that survives the next two quarters, Kimi K3.1 Pro at CNY 13 / CNY 68 is the current K-generation entry point with 1,048,576 tokens of context - but note it is not the cheapest way to serve that volume.
Frequently Asked Questions
Which Chinese LLM has the longest context window right now?
DeepSeek V4 Pro, Zhipu GLM-5.3, Alibaba Cloud Qwen3.8-Max (991K), Moonshot Kimi K3 and Kimi K3.1 Pro, and ByteDance Doubao-Seed-2.1-Pro (1024K) all sit at 1,048,576 tokens. DeepSeek V4 Pro also publishes the highest maximum output in the field at 384K tokens. Tencent Hunyuan Hy3 is the exception among domestic vendors with 256K on its API tier.
Which is cheaper per million tokens, Qwen3.8-Flash or GLM-5.3-Flash?
At list price they are tied at CNY 0.80 input with 1M context each, and Qwen3.8-Flash is marginally cheaper on output at CNY 2.70 versus CNY 2.80. Zhipu is currently running GLM-5.3-Flash at a limited-time 50% discount, so the real GLM rate right now is CNY 0.40 input and CNY 1.40 output - but the official page publishes no explicit end date, so treat it as subject to change. On cache hits the ranking flips back: Qwen3.8-Flash charges CNY 0.10 against GLM-5.3-Flash's CNY 0.23, or CNY 0.115 while the discount runs.
Is Kimi K2.7 still worth using, and what replaced Xiaomi MiMo-V2.5?
Kimi K2.7 should be treated as a legacy tier. Moonshot stopped opening it to new users on October 31, 2026 and plans to retire the whole K2 family by the end of December, and its 262,144-token context is a quarter of the 1,048,576 tokens available on K3 and K3.1 Pro. Migrate to K3 or K3.1 Pro. Xiaomi retired MiMo-V2.5 and MiMo-V2.5-Pro on October 21 at 10:00 Beijing time; the current generation is MiMo-V2.6, with MiMo-V2.6-Pro at CNY 3.00 / CNY 6.00 and CNY 0.025 cached input, and MiMo-V2.6-Flash at CNY 1.00 / CNY 2.00.
Why does cached input price matter more than headline price for agent workloads?
Because an agent loop re-reads the same context on every turn. With cached input at CNY 0.10 on Qwen3.8-Flash or CNY 0.025 on MiMo-V2.6-Pro, a hundred iterations over a 100K-token codebase cost a fraction of a single uncached pass; at the uncached rate the same loop is two orders of magnitude more expensive. Two details matter: Kimi bills cache writes separately by TTL (CNY 13.00 for five minutes, CNY 26.00 for one hour) while a hit is charged only at the hit rate, and Batch API halves the standard rate on both Zhipu and Alibaba Cloud, so offline processing should never run at the interactive price.
Want more AI pricing comparisons?
Check our price changelog for daily updates on AI model pricing, new releases, and deprecations.