How DeepSeek‑R1 Is Challenging OpenAI’s o1 and Shaping the AI Landscape

Benchmark Performance

On 24 January 2024 DeepSeek‑R1 achieved a score of 1357 on the Arena benchmark, ranking third overall among all model categories. In the StyleCtrl (style‑control) track it tied for first place with OpenAI o1, which scored 1352.

Model Architecture and Training

DeepSeek‑R1 is a large‑scale inference‑oriented model that relies on a limited amount of manually labeled data combined with extensive reinforcement learning from human feedback (RLHF). This training pipeline improves reasoning speed and accuracy on multi‑task benchmarks, including mathematics, code generation, and natural‑language reasoning.

Open‑Source Release

The full model and its distilled variants (e.g., DeepSeek‑R1‑Zero) are released under the MIT License. Distillation reduces model size while preserving performance comparable to OpenAI o1‑mini across several evaluation domains, making the models suitable for resource‑constrained environments.

API Access and Pricing

An HTTP API provides chain‑of‑thought generation. Pricing is tiered based on cache usage: 1 CNY per million input tokens when cache hits occur, up to 4 CNY per million input tokens without cache, and a flat rate of 16 CNY per million output tokens.

Implications for the Community

By open‑sourcing the model and offering a low‑cost API, DeepSeek lowers the barrier for researchers and developers to experiment with high‑performance reasoning models, encouraging broader adoption and collaborative advancement in the AI field.