The Chinese AI tech brand – DeepSeek has now officially introduced its new DeepSeek-V3.1 model, and it has been made available to users. With the DeepSeek-V3.1 model, the brand is taking its first step into the agent era.
Read more about it below.
DeepSeek-V3.1 – Introduced With Think and Non-Think Modes
As mentioned, DeepSeek has introduced its DeepSeek-V3.1 model as the first step the brand is taking into the agent era and the new model is revealed to offer two modes: Think and Non-Think. Notably, relying on DeepSeek-V3.1 model’s Think mode, users can expect answers and responses in a lesser time when compared to the DeepSeek-R1-0528 model. Additionally, it features stronger agent skills as the post-training has boosted tool use and multi-step agent tasks.
On the API update side, DeepSeek chat will now rely on the Non-Thinking mode, while DeepSeek reasoning will make use of the Thinking mode, and 128K context is available on both. Also, the Anthropic API format is supported, as well as the Strict Function Calling support (in Beta API).
Speaking of the upgrades with respect to tools and agents, on SWE (Verified, Multilingual) and Terminal benchmarks, the DeepSeek-V3.1 model achieved better results. For complex search tasks, it also offers a stronger multi-step reasoning capability. About the thinking efficiency, DeepSeek-V3.1 has attained improved results than the previous DeepSeek-R1-0528 model.
Coming to the model update, DeepSeek V3.1 base has been pre-trained with 840 billion tokens of data, and its ability to understand and process long context extensions/texts has been significantly improved. Furthermore, the tokenizer and chat template have been updated as well.
Along with the new DeepSeek-V3.1 model, the brand has also announced new pricing too, and the price change will be taking effect on the 5th of September 2025 (at 4:00 PM UTC). The input API price will be made USD 0.07 per 1M tokens (cache hit)/USD 0.56 per 1M tockens (cache miss), whereas the price of output API is going to be 1.68 per 1M tokens.
