DeepSeek V3 Outperforms GPT-4o on Coding and Math — At a Fraction of the Cost
China’s DeepSeek has achieved what many Western analysts dismissed as impossible: a frontier-grade language model that beats OpenAI’s flagship GPT-4o on coding and mathematical reasoning benchmarks, at roughly one-tenth the inference cost.
DeepSeek V3, the company’s latest dense transformer model, scores above GPT-4o on HumanEval (coding), MATH-500 (mathematics), and MMLU (general knowledge). The model uses a Mixture-of-Experts architecture that activates only 37 billion parameters per forward pass despite having 671 billion total parameters — making it both powerful and efficient.
The Price Gap Is Staggering
At $0.27 per million input tokens and $1.10 per million output tokens, DeepSeek V3 costs approximately 10 times less than GPT-4o ($2.50 input / $10.00 output). For organizations running high-volume AI workloads, this translates to orders-of-magnitude savings.
This cost advantage has forced OpenAI and Anthropic to reconsider their pricing strategies, with both companies introducing cheaper model tiers in response.
Open Source and Self-Hostable
Unlike its American counterparts, DeepSeek V3 is fully open-source under a permissive license, allowing organizations to self-host the model on their own infrastructure. This eliminates vendor lock-in and data privacy concerns — particularly relevant for governments and enterprises in the Global South.
The model weights are available on Hugging Face and can be run on consumer-grade clusters with quantization.
Geopolitical Dimension
DeepSeek’s success is emblematic of China’s broader strategy in AI: instead of competing on compute (where US export controls create disadvantages), Chinese labs compete on algorithmic efficiency — doing more with less.
The result is a model that challenges the narrative that frontier AI requires massive NVIDIA GPU clusters. DeepSeek reportedly trained V3 for under $6 million — a figure that shocked the industry when disclosed.
With information from DeepSeek research team.