DeepSeek V4-Pro is priced at $3.48 per million tokens—1/22 of GPT-5.5. We see the LLM war's main battlefield shifting from "who's strongest" to "who's cheaper and good enough."
What This Is
DeepSeek released V4 on the same day as GPT-5.5—no press conference, just a technical report and open source. Two models: V4-Pro (flagship, 1.6 trillion total parameters) and V4-Flash (budget, 284B total parameters), with the latter priced at just $0.28/M tokens.
Cheap doesn't mean burning cash on subsidies. V4 uses three tactics to crush costs: MoE architecture (Mixture of Experts, activating only 3% of parameters per inference), hybrid attention mechanism (under 1M token context, inference compute needs only 27% of the previous generation), and a three-tier inference mode (not every request needs "deep thinking").
The performance positioning is honest: lagging behind SOTA closed-source models by about 3-6 months. But it has bright spots in long-context understanding (MRCR 1M surpasses Gemini 3.1 Pro) and math (perfect score in the Putnam Competition). Migration cost is nearly zero—just change one line of code.
Industry View
The indie developer community is visibly thrilled. A real Reddit case: a 24/7 autonomous Agent system consumed 100M tokens over 4 weeks; the DeepSeek bill was $280, while GPT-5.5 would cost $3000+. This isn't saving pocket money; it's the difference between "can run" and "can't afford." DeepSeek also shows better consistency in long Agent loops, unlike models over-RLHF'd (Reinforcement Learning from Human Feedback) that easily "drift"—constantly apologizing, repeatedly confirming, and refusing to execute.
But we must question the sustainability of this low-price strategy. A 3-6 month performance gap is not small on the fast-iterating LLM track; for enterprise scenarios demanding extreme accuracy (healthcare, legal, finance), "good enough" might not suffice. Another view holds that V4's low price could squeeze the entire industry's profit margins, forcing rivals to follow suit, ultimately leaving no one with healthy profits—we've seen this script in the tech industry before.
Impact on Regular People
For Enterprise IT: LLM procurement costs have dropped by an order of magnitude, but we need to reassess the boundary of "good enough"—cheap doesn't mean suitable for all scenarios, especially core enterprise businesses with low fault tolerance.
For Individual Careers: Indie developers and small teams can now run 24/7 Agents (AI systems that autonomously perceive environments, make decisions, and execute actions). The barrier to AI application entrepreneurship shifts from "fundraising ability" to "product judgment."
For the Consumer Market: The foundational model price war means terminal AI applications have more room to cut prices. It's highly probable that consumers will get cheaper or even free AI tools.