DeepSeek just released V4.1-Flash, and it isn't a routine model bump. It's an architecture change specifically aimed at inference cost — aggressive enough that DeepSeek is retiring its own flagship, V4-Pro, in favor of it.
Starting September 14, every request sent to V4-Pro gets silently routed to V4.1-Flash at Flash pricing, and stays there until a V4.1-Pro exists. DeepSeek's own testing puts the smaller model ahead of the larger flagship on performance, cost, speed, and total runtime. On Terminal-Bench 2.1, it edges out Claude Opus 5 and GPT-5.6.
The cache is the story
The headline benchmark numbers are good. The KV cache numbers are the part that actually matters for anyone serving this model at scale. V4.1-Flash stores cache entries in 4-bit floating point, bringing the global footprint to 890 bytes per token — about a quarter of the previous generation, and more than 400x smaller than DeepSeek's own V1. Persistent SSD cache storage drops to roughly an eighth of what it needed before.
That matters because cache-hit costs are often the single largest line item in agentic workloads. Every multi-turn session, every long-context tool call, is repeatedly re-reading cache rather than recomputing from scratch. Shrink the cache four times over and you haven't just made a faster model — you've changed the unit economics of every agent built on top of it.
Competing on architecture, not just capability
The engineering underneath this is a new causal encoder-decoder split: a 552-billion-parameter mixture-of-experts model that keeps just 8 billion parameters active during prefill and 16 billion during generation, with a hierarchical sparse attention scheme (DeepSeek calls it CSA2) that decides, layer by layer, whether to fully recompute attention, reuse it from an earlier layer, or skip it entirely.
The pattern to watch is that DeepSeek isn't competing on raw capability alone anymore. It's competing on cost-per-token-served at the architecture level — and forcing every other provider serving comparable capability to answer the same question: what does your cache actually cost you?