After a span of about half a year, DeepSeek has finally released a new version of its model. This model is also open-source, and you can currently use it on DeepSeek's official website and API platforms. It is worth noting that the new release splits into two distinct models: V4 Flash and V4 Pro. There is a massive difference in their parameters, as well as a significant difference in their API costs.
The "Quick Mode" on the official website utilizes the Flash model, while the "Expert Mode" uses the Pro model. The performance gap between the Pro model and Claude Opus 4, GPT 5.4, and Gemini 3.1 Pro is not significant. It has also widened the gap with other domestic Chinese models to a certain extent, and DeepSeek continues to offer lower prices.
The API pricing is shown in the table below:
It is worth mentioning that Flash still maintains its consistently ultra-low cost, but Pro is relatively more expensive. After all, it has significantly more parameters and requires a much higher computational load.
Comments NOTHING