DeepSeek-R1 is open, and cheap to call

On 20 January 2025 DeepSeek released R1. It says large-scale reinforcement learning in post-training brought math, code, and reasoning in line with OpenAI’s o1. Weights, code, and a technical report went out under the MIT License. Six distilled models (Llama and Qwen bases) were open-sourced with it.

The API model name is deepseek-reasoner. The lab listed $0.14 / $0.55 per million input tokens (cache hit / miss) and $2.19 per million output tokens. Chat is on chat.deepseek.com with DeepThink turned on.

Open and cheap changes who can try a reasoning model. It does not settle whether a given proof, patch, or citation is right. Check the paper and the output.