DeepSeek-V4-Pro-0813 is officially launched on SiliconFlow, and it's a Day-0 launch. The core highlights of this update are three numbers and one orientation: a 1M context window, three levels of reasoning intensity (low/medium/high), and a stronger focus on coding, tool calls, and agent workflows. The open-source license remains unchanged, still MIT.
The pricing structure is clear: $1.32 per million input tokens, $3.96 per million output tokens, and only $0.44 for cache-hit tokens. The last figure is truly worth noting—the price of cache-hit tokens being this low means that for Agent workflows that rely heavily on repeated calls and long context inputs, the unit cost calculation in long-text scenarios now has a more concrete anchor point, avoiding budget overruns when handling long contexts.
