Chinese AI models undercut OpenAI and Anthropic on price but raise enterprise risk
DeepSeek's API costs $0.14/million tokens, 30x cheaper than GPT-5.5 — but at what cost?
Chinese AI models like DeepSeek V4 Flash and V4 Pro are undercutting OpenAI and Anthropic on API pricing by a wide margin — DeepSeek V4 Flash costs $0.14 per million input tokens compared to OpenAI GPT-5.5 at $5 and Claude Sonnet 4.6 at $3. For volume-heavy workloads (coding, summarization, internal search), this can yield millions in savings, as demonstrated by Lindy AI, which moved some workloads from Anthropic to DeepSeek. The lower cost has made Chinese models prominent on platforms like OpenRouter, and the narrowing performance gap between open and closed models (per Stanford HAI’s 2025 AI Index) further fuels adoption.
Yet enterprise adoption comes with a longer checklist. The core trade-off is between cost and risk: data governance, security review, model behavior, compliance, and vendor accountability. Self-hosting open-weight models can keep data inside company infrastructure, but opens abuse risks as agentic AI systems call tools and modify code. Access through US cloud intermediaries can reduce data-routing concerns but may limit direct vendor accountability. Direct use of China-hosted services carries the highest exposure, especially for regulated workloads. Procurement teams must ask which models are running, where company data goes, who operates the infrastructure, and what law governs the provider.
- DeepSeek V4 Flash costs $0.14/million input tokens vs OpenAI GPT-5.5 at $5 and Anthropic Claude Sonnet 4.6 at $3.
- Lindy AI, a San Francisco startup, saved millions by moving some workloads from Anthropic to DeepSeek via a US provider.
- Security researchers warn open-weight Chinese models can be adapted for malicious use, raising procurement and compliance concerns.
Why It Matters
Teams can cut AI costs dramatically by routing non-sensitive tasks to Chinese models — but must balance savings against data governance and vendor risk.