DeepSeek-R1 Upsets AI Market With Low Prices

Chinese company DeepSeek disrupted the AI market with the launch of its R1 model last week. As a result, the Nasdaq fell by 3 percent over the course of Monday, driven by losses of chip maker Nvidia of nearly 17 percent. A research paper published along with the launch said that DeepSeek spent only $6 million on computing power per training run for the model, below the costs that are estimated (and have been partially confirmed) for popular AIs like ChatGPT or Google's Gemini. This is despite the fact that since going live on January 20, DeepSeek-R1 has earned good marks for its performance that rivals its larger competitors. High-end computer chips, like the ones produced by Nvidia, are central to developing and running large AI models, but this new development now shows that quality results in AI can also be achieved with a smaller budget, fewer chips or less advanced ones. 

The lower price associated with DeepSeek-R1 is also visible in pricing info published by chatbot provider DocsBot. The price of using an AI model commercially is usually broken down into the use of tokens (the smallest AI model processing unit, around 4 characters long). Uploading 1 million tokens into DeepSeek-R1 costs just 55 cents, according to the DocsBot website. Downloading 1 million tokens cost $2.19, according to the latest data. The most comparable AI model to R1 by U.S. company OpenAI is ChatGPT-o1 Mini, a toned down version of its latest o1 AI model that is capable of text processing only instead of text and image that are available in the full version. Like DeepSeek-R1, ChatGPT-o1 Mini has the ability to use the input of 124,000 tokens as context for a conversation (before forgetting the earliest prompts). While DeepSeek's program can output 32,000 tokens in a single request, ChatGPT-o1 Mini can put out a maximum of 65,500. The grades the two programs receive on a number of industry benchmarks that test subject knowledge, understanding, reasoning capabilities, accuracy and consistency are comparable however, with DeepSeek-R1 even outperforming its rival slightly.

AI model Grok by Elon Musk's company xAI also boasts a larger output window (128,000 tokens per request) as well as being text and image capable, but also slightly underperforms on benchmark tests in comparison to R1. Both U.S. models are much more expensive than R1 in a commercial use scenario, costing $3-$5 for 1 million token inputs and $12-$15 for 1 million outputs.

Google's Gemini and Amazon's Nova AIs are cheaper than OpenAi's and xAI's products, but still slightly more expensive than R1. However, they are only partially comparable to the other programs on this list as they were developed to handle input of multimedia formats, therefore sporting much larger input windows (while having smaller output windows - potentially because they were tailor-made for customer and search support functions instead of the broader LLM approach other models took). In the performance tests, the models are once again performing similarly to DeepSeek-R1. Finally, an even cheaper option based on open-source technology is Nvidia's Llama 3.1 Nemotron 70B Instruct. The text-only model uses Meta's Llama AI and has received good marks from users as well as on performance tests.