What changed
On September 10, 2026, DeepSeek released V4.1 Flash, a 552B MoE with 8B active input and 16B active output parameters, native vision, new architecture and lower prices. Legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp aliases route to V4.1 Flash. The original release announcement also said V4 Pro would reroute to Flash from September 14, but the current official API Change Log says V4 Pro service continues with unchanged billing, and the live pricing page lists V4-Pro-0813 separately. The two primary documents contradict one another on Pro; do not assume the Pro reroute happened.
Why it matters
The Flash migration is real for users of the old Flash aliases: their backend model and pricing have changed. Builders should re-benchmark behavior, tool use, vision, latency and cost. But teams using V4 Pro should not conflate the confirmed Flash alias migration with an unverified Pro migration: current provider documentation shows Pro continuing separately.
V4.1 Flash replaces the old split between text Flash and Vision-Exp
DeepSeek V4.1 Flash now combines native multimodal understanding with the Flash family’s text and agent use cases. DeepSeek says the older V4 Flash and V4 Flash Vision experimental compatibility names temporarily route to the new model, so existing integrations may receive different model behavior without a model-string change.
The new architecture is designed around cheaper serving
V4.1 Flash is a 552B MoE model, but DeepSeek says only 8B parameters are active for input and 16B for output. The company also says its KV cache needs one quarter of the HBM and one eighth of the SSD storage required by the prior generation. Those are vendor claims, but they explain the lower pricing and higher-throughput positioning.
Pricing drops while time-of-day pricing remains
DeepSeek says V4.1 Flash launched with lower API rates on September 10 and retains peak/off-peak pricing, with off-peak rates at half peak rates. Builders with deferrable workloads can still combine model choice, caching and scheduling as cost-control levers.
V4 Pro remains listed separately despite the older reroute announcement
The September 10 release announcement says Pro requests would move to Flash from September 14. DeepSeek's current API Change Log instead says it decided to continue Pro service with unchanged billing; the live Models & Pricing page lists deepseek-v4-pro as V4-Pro-0813 with distinct rates. Without authenticated request traces, the safe statement is that the official documents conflict and current operational docs retain Pro.
Existing users should treat this as a backend migration
Because compatibility model names can point to a new model family, regression testing should cover instruction following, tool use, multimodal behavior, latency, token consumption and task success. The safe assumption is that a stable model string does not guarantee a stable underlying model.