DeepSeek has made an announcement regarding the upgrade of its online model to the DeepSeek-V3.1-Terminus version. This iteration presents two distinct modes: the thinking mode and the non-thinking mode. Both modes boast a substantial context window of 128k, which is now accessible for online users to experience. Specifically, the deepseek-chat feature aligns with the non-thinking mode, whereas the deepseek-reasoner feature embodies the thinking mode.

This upgrade not only retains the existing functionalities but also integrates enhancements derived from user feedback. These improvements encompass the mitigation of language consistency challenges and the refinement of Agent capabilities. By default, the output length for the non-thinking model is configured at 4K, with the flexibility to extend up to 8K. Conversely, the thinking model's output length is preset at 32K, with the capacity to reach a maximum of 64K.

Regarding pricing, when the input cache is successfully utilized, the cost amounts to 0.5 yuan per 1 million tokens. In scenarios where the input cache is not employed, the cost rises to 4 yuan. Additionally, the output price stands at 12 yuan.