Two DeepSeek models and Meta's flagship dense model — each targeting a different point on the capability-cost frontier. DeepSeek R1 is a reasoning-specialized model trained with reinforcement learning to generate explicit chain-of-thought traces before producing final answers. On GPQA Diamond and competition math it outperforms much larger dense models. The chain-of-thought process adds output tokens, which increases both latency and per-query cost, so the premium is appropriate only for tasks where reasoning-trace quality matters — formal proofs, multi-hop scientific QA, or workflows where auditability of the reasoning path is a requirement. Context window is 131K. DeepSeek's commercial license terms need verification before deployment. DeepSeek V3.2 is the May 2025 successor to V3, a mixture-of-experts model with roughly 37B active parameters per forward pass and a ~30% inference-cost reduction over V3. On code, math, and general reasoning benchmarks it delivers performance well above what its inference cost implies, with a 131K context window and broad provider availability. Where R1 optimizes for explicit reasoning depth, V3.2 optimizes for cost-efficient general capability across a broad task surface. Same commercial license caveat applies. Llama 3.1 405B Instruct at 405B dense parameters offers the broadest knowledge coverage of the three — MMLU scores near the top of open-weights models at its July 2024 release, strong general-instruction following, 131K context, and the Llama 3 community license for commercial use. Per-token cost is highest in this group due to multi-GPU serving requirements. Pick DeepSeek R1 when chain-of-thought reasoning quality on GPQA-class tasks is the evaluating criterion. Pick DeepSeek V3.2 for strong general performance at the best cost-efficiency ratio. Pick Llama 3.1 405B when licensing flexibility to self-host and broad knowledge coverage are the priority.