AI
Apr 10, 2026DeepSeek-V4 keeps the Chinese frontier 6× cheaper than Western models
DeepSeek-V4 matches GPT-5.4 on most benchmarks at $0.27 per million input tokens. The pricing gap with the West is now structural, not promotional.
DeepSeek released V4 last week with a pricing card that makes the gap with Western frontier models impossible to ignore: $0.27 per million input tokens, $1.10 per million output. GPT-5.4, the closest Western analogue on benchmarks, lists at $1.50 / $10. That is a 5.5–9× cost difference for comparable output quality on most evals.
The quality story is real. DeepSeek-V4 trades blows with GPT-5.4 across MMLU-Pro, HumanEval, and IFEval, edges ahead on math, and trails on multi-step agentic tasks. For high-volume inference—support automation, content pipelines, classification—the cost gap dominates the quality gap.
The pricing isn’t a loss-leader. DeepSeek’s technical report attributes the cost structure to MoE sparsity, FP8 training, and a custom training stack on H800-tier hardware. This is the price floor now, not a promotion. Western labs that depended on token margins for funding R&D will need a new economic story.
Source
deepseek.com