Model positioning and release date
DeepSeek-R1 is DeepSeek's first-generation reasoning model, intended for tasks such as mathematics, programming, and natural language reasoning. The model was officially released on January 20, 2025, with model weights made available and an API launched at the same time; the date recorded here is the official initial release date, not a placeholder date at the end of the month. (api-docs.deepseek.com)
Capabilities and uses
- Large-scale reinforcement learning was used during post-training. The model can be used for complex problem-solving, coding assistance, and research; the initial API provided chain-of-thought output. DeepSeek states that its performance on mathematics, coding, and reasoning tasks is comparable to OpenAI o1; this is the vendor's benchmark conclusion and does not mean the models are equivalent on all tasks. (api-docs.deepseek.com)
- The DeepSeek-R1 weights listed in the official repository have 671B total parameters, 37B activated parameters, and a 128K context; these are specifications for the model weights and should not be directly treated as specifications for the current API. DeepSeek states that the model weights use the MIT license. (github.com)
Limitations and API versions
The model may produce hallucinations or incoherent or repetitive output; important results should be independently verified. A subsequent official update stated that the newer R1 reduced hallucinations compared with the older version; its repository also provides usage recommendations concerning repetitive or incoherent output. (api-docs.deepseek.com)
deepseek-reasoner was the API name for calling R1 at launch; it is not a version identifier permanently bound to the initial weights. Official records state that this name was upgraded to DeepSeek-R1-0528 in May 2025 and subsequently pointed to other models; an April 2026 announcement stated that it then pointed to V4-Flash's thinking mode and was planned to be discontinued on July 24, 2026. (api-docs.deepseek.com)
Source note: In the current data, whether deepseek-reasoner was actually discontinued as planned on July 24, 2026 has not been independently verified; that date is based on the official discontinuation plan, not confirmation of an actual discontinuation event. Neither the current Chinese nor English official pricing page lists that name or the initial R1, so they do not establish that it currently still provides the initial R1 service. (api-docs.deepseek.com)
Official API pricing
The following are historical prices from the 2025 initial release announcements, not currently available prices; the retrieval date for all prices was November 20, 2026. The two currencies come from official announcements in different languages; no exchange-rate conversion is made, and they are not treated as the same plan. (api-docs.deepseek.com)
- Domestic official announcement, Chinese yuan CNY, per million tokens: input cache hit 1 yuan, input cache miss 4 yuan, output 16 yuan. Domestic official release and pricing announcement. (api-docs.deepseek.com)
- International official announcement, US dollars USD, per million tokens: input cache hit $0.14, input cache miss $0.55, output $2.19. International official release and pricing announcement. (api-docs.deepseek.com)
Source note: In the current data, the current official API pricing for the initial release of DeepSeek-R1 has not been independently verified; the amounts above can only be confirmed as prices stated in the initial release announcements. The current domestic pricing page and the current international pricing page list other models, and their prices cannot be applied to the initial R1. (api-docs.deepseek.com)