Overview
- DeepSeek opened a limited internal test of a time‑bound V4.1 Flash endpoint, deepseek-v4.1-flash-expires-on-0910, that developers can call with the existing base_url, is capped at 20 concurrent connections per account, and is billed like the current V4 Flash.
- The platform announced a Flash‑series price cut that takes effect on Sept. 10 at 12:00 Beijing time, setting idle prices at 0.02 RMB per million tokens for cache‑hit input, 1 RMB per million for cache‑miss input, and 4 RMB per million for output with peak periods charged at double those rates.
- DeepSeek says V4.1 Flash—described by the company as using a new model architecture with native multimodal support—outperforms the existing V4 Pro on internal and external tests and will be used to route V4 Pro requests and bill them at Flash prices during the transition.
- The firm has merged its Fast, Expert and Vision modes into a single intelligent mode that auto‑detects task complexity and activates visual processing when images are present, but DeepSeek has not yet published technical details, parameter counts or independent benchmarks for V4.1 Flash.
- User savings will vary: customers with high cache‑hit rates, small output volumes or flexible schedules can see the largest cuts, while heavy output or peak‑time workloads may see smaller gains; the change could widen Flash adoption and ease DeepSeek’s inference‑cost pressure if the claimed performance holds up in production.