DeepSeek V4.1 Flash Launches

DeepSeek emailed users this afternoon saying the new V4.1 Flash beats its own V4 Pro flagship on performance, cost, speed, and total runtime. V4 Pro will be retired in 4 days, with all requests rerouted to Flash at the cheaper price.

💡 What You Will Learn

DeepSeek emailed users this afternoon saying the new V4.1 Flash beats its own V4 Pro flagship on performance, cost, speed, and total runtime. V4 Pro will be retired in 4 days, with all requests rerout

V4 Pro only lived 28 days (Aug 13 release → Sep 14 retirement). Between V4.1 Flash and the next V4.1 Pro, the flagship slot is empty. DeepSeek isn't just stacking parameters — V4.1 uses a new architecture with native multimodal support (vision folded into the main model instead of a separate vision module), and the heavy lift is in post-training rather than model size. Community tests report 300+ tokens/sec. The real lever, though, is cache hits: at 0.02 RMB per million cached tokens, an Agent run with 99% cache hit rate on ~100M tokens can cost under 1 RMB. Concurrency tells the real story — V4 Pro caps at 500 concurrent requests, Flash at 2,500. The reason the flagship slot is left empty on purpose is that the heavy, expensive flagship fights against DeepSeek's 'cheap model' narrative; the rumored V4.1 Pro is being held back to match Anthropic's Fable 5.1. Cold water: 'comprehensive superiority' is DeepSeek's self-measured benchmark — third-party independent benchmarks haven't caught up. Beta users report the new model over-thinks, burning extra tokens in chain-of-thought. Heavy users report 53 RMB per day bills. Cheap is cheap against overseas flagships, not free.