DeepSeek makes 75% V4-Pro price cut permanent, repricing intelligence-per-dollar

DeepSeek's announcement that the 75% discount on V4-Pro is now permanent — confirmed via its official @deepseek_ai account ('We are making our discount permanent! 🎉', 23K likes) — turns what looked like a promo into structural pricing. API rates of 0.025 to 6 yuan per million tokens are well under Western competitors and explicitly aimed at developers and enterprise customers inside China's hyper-competitive AI market.

Mechanically, the move is enabled by two things: aggressive use of domestic chip alternatives (reducing dependence on NVIDIA hardware constrained by export controls) and DeepSeek's well-documented MoE training efficiencies. The pricing also lands in the same week Microsoft canceled Claude Code seats over token cost and Google pitched $1B in Gemini enterprise savings — DeepSeek isn't fighting the price war, it's setting the floor.

Competitive context: Reuters and Bloomberg both confirmed the permanent cut, and developer forums are spreading benchmark comparisons that put V4-Pro near GPT-5.5 quality at a small fraction of the cost. Alibaba's Qwen3.7-Max release (35 hours of autonomous kernel optimization on a custom chip) reinforces that Chinese labs are simultaneously pushing capability and price aggressively. Meanwhile US-side users complain about Google and Anthropic 'nerfing' model quality — a perception that gives DeepSeek an opening for international adoption where regulation permits.

What to watch: whether US frontier labs match with structural price cuts (vs. promo credits), whether Western enterprises actually deploy V4-Pro given data-residency concerns, and whether US export controls expand to limit DeepSeek's cloud distribution. The intelligence-per-dollar curve just shifted, and every CFO running an LLM workload is going to ask why their inference bill isn't moving with it.