The Chinese artificial intelligence sector has been thrown into an aggressive, all-out pricing war, ignited by a series of hyper-competitive moves from startup DeepSeek.

By drastically undercutting historical market rates, DeepSeek has forced China’s biggest tech conglomerates and AI laboratories into a race to the bottom, fundamentally changing the economics of AI deployment.

The Catalyst: DeepSeek Makes a Massive 75% Cut Permanent

In late May 2026, DeepSeek sent shockwaves through the global tech industry by announcing that its massive 75% promotional discount on its flagship model, DeepSeek V4-Pro, would become permanent.

• The Disruptive Price Point: The permanent adjustment fixes DeepSeek V4-Pro at just $0.435 per million input tokens (and an industry-low $0.003625 for cache hits) and $0.87 per million output tokens.
• The Global Contrast: This pricing effectively positions DeepSeek at roughly 34 times cheaper than Western frontier benchmarks like GPT-5.5 and 17 times cheaper than Claude Opus 4.7.
• Efficiency Over Subsidization: DeepSeek has clarified that this is not a venture-backed cash burn or temporary marketing stunt. The V4-Pro architecture was engineered natively to reduce long-context memory and compute footprints. Furthermore, the model has been fully optimized to run on domestic Huawei Ascend 950 silicon instead of restricted American Nvidia hardware, allowing DeepSeek to dramatically slash infrastructure costs.

The Domino Effect: How Rivals Responded

DeepSeek’s aggressive pricing structure gave domestic rivals a tough ultimatum: drop prices immediately or face mass developer churn. The market immediately fragmented into two core survival strategies:

1. The Volume Race (Slashing to Zero)

• Xiaomi (MiMo): In an eye-watering retaliation, Xiaomi slashed the API costs for its MiMo-V2.5 model by 99%, permanently flattening its long-context pricing structure to a flat $1 input / $3 output per million tokens. The volume play paid off instantly, with MiMo processing an astronomical 1.7 trillion tokens in a single week.
• Tencent Cloud: Tencent took an even sharper scalpel to its platform, axing the costs of hosting DeepSeek-V4 series models on its intelligent agent development platform by up to 97.5% to protect its enterprise ecosystem.
• Alibaba (Qwen) & Moonshot (Kimi): Alibaba stabilized its flagship Qwen3 Max at a competitive $3.90 per million output tokens, while Moonshot’s Kimi K2.6 began competing heavily on long system prompts by dropping its cache-hit floor to a microscopic $0.07 per million tokens.

2. The Hybrid Pivot (Refusing to Bleed)

• MiniMax: Recognizing that racing to zero on per-token pricing is structurally unsustainable for smaller firms, startup MiniMax went the opposite direction. With the launch of its M3 model, it abandoned pure consumption billing in favor of a hybrid system, pairing baseline token access with monthly subscriptions ranging from $7.24 to $69.28.

Strategic Comparison of Top Chinese LLM APIs

The Macro View: What This Means for Global AI

The pricing bloodbath in China is triggering a major shift in how the corporate world views the return on investment (ROI) of artificial intelligence.

The presence of high-performing, ultra-cheap models is making premium, high-margin consumption pricing from Western giants much harder to justify for standard, repeatable enterprise workloads. Industry analysts anticipate that this pricing pressure will ultimately force premium global labs to transition away from basic per-token fees toward more defensible, value-based subscription and enterprise outcome models.