DeepSeek's V4 Flash Refreshes the Bottom of the Price Curve as the Cheap-Tier Model War Accelerates

DeepSeek's V4 Flash 0731 build has taken the price-efficiency lead in several real-task evaluations, the latest move in a segment that has quietly become the most strategically important in the model market: the cheap workhorse tier. The competitive logic is unchanged from the summer's earlier rounds — Google's Gemini 3.6 Flash and Alibaba's Qwen3.8 Max target the same buyers — but the pace is accelerating, with industry trackers now counting over 337 model releases across major labs and the bottom of the price curve refreshing roughly monthly. Price-efficiency at the task level, not the token level, is the metric that matters here. A model that costs half as much per token but needs twice the tool calls to finish a job saves nothing; evaluations that measure completed real tasks per dollar consistently reshuffle the rankings that raw benchmark tables suggest. For operators of high-volume pipelines — classification, extraction, routing, summarization at scale — the practical consequence is that the optimal model choice now has a shelf life of weeks. Teams that hard-wire a single provider into their automation inherit whatever that provider's pricing does next; teams that keep an evaluation harness and a thin abstraction layer can re-point their volume to whichever cheap tier currently wins. That discipline — a benchmark you own, run against your own tasks — is worth more than any individual model release this cycle produces.