DeepSeek R1 — pricing, providers, and benchmarks

DeepSeek R1, released in January 2025 under the MIT license, was the first open-weights model to credibly match OpenAI's o1 series on reasoning benchmarks like AIME and MATH. It's a 671B-parameter MoE that uses extended chain-of-thought reasoning at inference time, which means it's more expensive per token than its non-reasoning siblings — typically $0.40–$0.55 per 1M input tokens and $2.00–$2.50 per 1M output tokens, since reasoning chains generate substantially more output tokens than a standard completion. Best for problems where verifiable correctness matters more than per-call cost: math, formal proofs, code synthesis with strict tests, and multi-step planning. For anything not requiring deep reasoning, V3.2 is a better cost choice.