DeepSeek launches V4-Pro-0813 model, only 0 is needed to output one million Tokens

DeepSeek launches the V4-Pro-0813 model, and the output of one million Tokens only costs 0.87 US dollars. As the competition among the major AI giants becomes increasingly fierce, the well-known Chinese artificial intelligence laboratory DeepSeek (Depth Quest) once again makes a big impact. Recently, DeepSeek officially launched its latest flagship model - DeepSeek-V4-Pro-0813, and simultaneously launched it on the DeepSeek API and DeepSeek Chat platforms. According to the official pricing information, the input Token price of the DeepSeek-V4-Pro-0813 model is US$0.435 per million, and the output Token price is US$0.87 per million. This move is seen as another strong move for DeepSeek to directly compete with industry giants such as OpenAI and Anthropic. According to preliminary test data leaked on social platforms, the model has 1.6 trillion total parameters, 49 billion activation parameters and 1 million context windows. Its performance in many mainstream benchmark tests such as Terminal Bench 2.1, Cybergym, DeepSWE and AutomationBench has surpassed Anthropic’s Opus 4.8. At the same time, the cost has been reduced by about 57 times, showing a very high cost performance. Not long ago, OpenAI significantly lowered the price of its GPT-5.6 Luna model, trying to start a price war with discounts of up to 80%. However, DeepSeek quickly launched the V4-Flash-0731 model with 284 billion parameters. It completely counterattacked at an extremely low price of only US$0.14 per million input tokens and only US$0.28 per million output tokens, and demonstrated powerful performance comparable to multi-trillion parameter models. Third-party industry statistics show that the total token consumption of DeepSeek ranked second in the world in July this year, second only to Anthropic. With the intensive release and popularization of new models of the V4 series, outsiders predict that its Token usage in August this year may even reach the top of the industry. Due to the explosive growth in call demand and the fact that the laboratory currently only deploys a computing cluster of about 20,000 NVIDIA H100 GPUs, some users have reported that their inference speed has slowed down during peak periods. How to maintain the ultimate cost-effectiveness while coping with the huge pressure of computing power will become the core challenge that DeepSeek will face next. via cnBeta.COM - Chinese industry information station (author: Source: cnBeta.CO