This is a genuine reasoning-versus-scale comparison, with three different architectures hitting different points on the benchmark-to-cost curve. DeepSeek R1 is a reasoning-specialized MoE model trained with reinforcement learning to produce explicit chain-of-thought traces before answering. On GPQA and competition math, it outperforms dense models several times its active-parameter count. The tradeoff is latency: reasoning traces add tokens, which increases time-to-first-response and total cost. Context window is 131K. DeepSeek's license terms require verification for commercial use. This model is the right choice when you need defensible step-by-step reasoning — scientific question answering, formal proofs, or tasks where auditability of the reasoning chain matters. Llama 3.1 405B Instruct is Meta's largest dense open-weights model from July 2024, 405 billion parameters, 131K context, and the broadest knowledge coverage in this group. MMLU scores are among the highest for any open model at launch. The cost is significant — multi-GPU serving at 405B means hosted pricing is substantially above 70B tiers, and provider availability is limited to those with large-memory inference infrastructure. For complex document synthesis, long-context analysis, or frontier reasoning without the explicit chain-of-thought style of R1, this is the reference. Qwen 3 72B Instruct from April 2025 is the practical workhorse at a fraction of 405B serving cost. It covers MMLU, HumanEval, and multilingual benchmarks solidly. What it lacks relative to R1 is explicit reasoning trace quality on hard GPQA-style tasks, and relative to 405B it lacks depth on knowledge-intensive questions. Pick DeepSeek R1 when chain-of-thought reasoning quality and GPQA/math performance are the primary evaluation criteria. Pick Llama 3.1 405B when raw frontier capability and broad knowledge coverage justify the cost. Pick Qwen 3 72B for strong all-around performance at an accessible price point.