Chinese AI lab DeepSeek has slashed API fees for its newly released flagship V4-Pro model by 75%, Bloomberg reported on April 27, 2026, in a move that reignites the Chinese AI price war and sharpens the cost gap between open-weight Chinese models and US frontier systems from OpenAI and Anthropic.
The limited-time discount, which runs through May 5, drops V4-Pro pricing to 3 yuan per million input tokens and 6 yuan per million output tokens. DeepSeek also cut input-cache-hit fees across its model family to roughly one-tenth of their previous rates — a significant break for developers with repetitive prompt patterns, agent workflows, or long shared context windows.
A direct shot at Western pricing
The new rates land far below comparable tiers from US labs. According to Bloomberg, output prices from Western competitors typically range from $12 to $25 per million tokens, putting V4-Pro's promotional pricing at a small fraction of that. Decrypt and Fortune both reported V4-Pro at near state-of-the-art quality for roughly one-sixth the cost of Anthropic's Claude Opus 4.7 and OpenAI's GPT-5.5, with some outlets describing the Pro tier as 98% cheaper than GPT-5.5 Pro for equivalent workloads.
DeepSeek released V4-Pro and V4-Flash in preview on April 24. V4-Pro reportedly carries around 1.6 trillion parameters, making it the largest open-weight model publicly available, and was trained with heavy use of Huawei's Ascend chip clusters and Supernode interconnect — part of China's broader push to reduce dependence on Nvidia silicon.
Price war as a strategy
The cuts mark the second major Chinese AI price war in roughly two years. Analysts cited by Bloomberg framed DeepSeek's move as implementing an aggressive "AI price reduction" posture even as compute, memory, and energy costs rise across the industry. The strategy appears designed to lock in developer mindshare during V4's preview window, before Alibaba's Qwen, Moonshot's Kimi, and Tencent's Hunyuan teams respond with their own discounts.
Not every analyst is convinced the economics work. A Startup Fortune analysis noted that V4-Pro reportedly costs roughly 15 times more to run than DeepSeek's prior V3.2 model on a per-request basis, complicating the narrative that each generation is automatically more efficient. Sustained discounts at this level may rely on either tolerating thin margins or front-loading user acquisition before raising prices.
Implications for the AI market
For enterprise buyers, V4-Pro's pricing reframes the make-or-buy calculus on agent workloads, where token volumes can dwarf the cost of a single user-facing query. For Western labs, it intensifies pressure to either match on cost — even as OpenAI moved in the opposite direction by roughly doubling per-token rates with GPT-5.5 — or to defend margins by leaning harder on differentiated capabilities like native computer use, longer-running autonomous agents, and verified safety guarantees.
For regulators and procurement officers, the discount may also accelerate scrutiny: cheaper Chinese open-weight models running on Huawei silicon remove two of the friction points that have so far slowed cross-border adoption. Whether DeepSeek can hold this pricing past May 5 will be the next data point worth watching.
