Key takeaways
- Launch DeepSeek-V3.1, a hybrid AI model, offering improved agentic task performance.
- Price DeepSeek-V3.1 at $0.56 and $1.66 per million tokens, making it cost-effective.
- Utilize 671 billion parameters for enhanced tool calling and thinking efficiency.
- Access the model via MIT License and download from Hugging Face.
- Activate thinking mode with a simple toggle in the DeepSeek Chat interface.
DeepSeek, the Chinese AI lab backed by High Flyer Capital, has announced a new AI model, called DeepSeek-V3.1. It is a hybrid model that supports both thinking (reasoning) and non-thinking modes. DeepSeek, in the announcement, also called it the first step towards the agent era.
The new models improve upon DeepSeek-V3 by providing better tool calling and enhanced thinking efficiency. They deliver answer quality comparable to the reasoning model DeepSeek-R1 while responding quickly. DeepSeek also stated that it has stronger skills in agent-based tasks, and post-training has increased tool use and multi-step task performance.
The model has a total of 671 billion parameters, with 37 billion active parameters. DeepSeek’s new models are available under the MIT License, and the models can be downloaded from Hugging Face. The models can also be accessed via DeepSeek Chat, and thinking can be activated via a toggle button.
DeepSeek also stated that it is priced at $0.56 and $1.66 per million input and output tokens, respectively, which takes effect from September 5.
Pricing Changes 💳
— DeepSeek (@deepseek_ai) August 21, 2025
🔹 New pricing starts & off-peak discounts end at Sep 5th, 2025, 16:00 (UTC Time)
🔹 Until then, APIs follow current pricing
📝 Pricing page: https://t.co/IyYitNzedg
5/5 pic.twitter.com/omxJiCBvGd
This is significantly cheaper than the newly announced GPT-5 model from OpenAI, which is priced at $1.25 and $10 per million input and output tokens, respectively.
In terms of performance, the DeepSeek-V3.1, in thinking mode, scored 80.1% on GPQA Diamond, which assesses it on PhD-level scientific reasoning questions. The latest version of the DeepSeek-R1 models scores 81%, while xAI’s Grok 4 leads the benchmark at 88%, followed by OpenAI’s GPT-5 (high) at 85%, and Gemini 2.5 Pro at 84%.
In the Aider Polyglot benchmark, which assesses the ability of AI models to follow instructions and edit code successfully without human intervention, DeepSeek-V3.1, with thinking, scored 76.3%. OpenAI’s GPT-5 leads the evaluation with 88%, o3-Pro scored 84.9% and Gemini 2.5 Pro scored 83.1%.
