Same lab, fundamentally different design objectives. DeepSeek R1 is a reinforcement-learning-trained reasoning model — it generates extended chain-of-thought traces and excels on tasks where explicit step-by-step derivation beats retrieval. [DeepSeek V3](/models/deepseek--deepseek-v3) is a 671B Mixture-of-Experts dense-context model optimized for throughput, general capability, and instruction following at scale, with ~37B active parameters per forward pass. The cost picture separates them clearly. DeepSeek V3 on commodity providers typically runs $0.14–$0.28/1M input tokens. DeepSeek R1 carries a premium — expect $0.50–$1.50/1M depending on provider — because the extended thinking overhead and GPU-hours per token are meaningfully higher. If you're running tens of millions of tokens per day, that delta compounds fast. [DeepSeek R1](/models/deepseek--deepseek-r1) is the right call for tasks where accuracy on hard reasoning problems justifies the cost: competitive math, multi-step code proofs, formal verification assistance, or research workflows where you're willing to pay for a model that shows its work and catches its own errors. Its AIME 2024 and MATH-500 scores are significantly above V3. DeepSeek V3 wins for high-volume general workloads: long-form drafting, RAG pipelines over large document sets, classification at scale, or agentic loops where most steps don't require deep reasoning — just reliable instruction execution at low per-token cost. Pick DeepSeek R1 if your task requires verified, step-by-step reasoning and budget allows. Pick DeepSeek V3 if throughput and cost efficiency matter more than reasoning depth.