What just happened

DeepSeek, the Chinese lab that shocked the AI world with a cheap frontier model in early 2025, has done it again. Its new model, V4.1 Flash, was built to fix one problem: AI agents are expensive to keep alive. DeepSeek says it cut the memory cost of running an agent to roughly one-eighth of what it was before.

Why agents are so expensive

An AI agent isn't a single answer. It's a worker that runs for minutes or hours, step after step. Ask it to research a company and it opens files, runs searches, checks its own work, fixes mistakes. Every step adds to what it must remember — and the longer it works, the more history it has to drag along. That memory costs real money every time the model takes another step.

How DeepSeek cut the cost

DeepSeek didn't just build a smaller model. V4.1 Flash is huge: 552 billion parameters, plus a separate 196-billion-parameter memory system called Engram, and a context window of up to one million tokens. Instead, it changed how the model remembers. Only half the model reads everything. Layers share notes instead of each keeping its own copy. Memories get compressed into a fraction of the space. Short-lived memories get dropped entirely. The result: one-quarter of the working memory and one-eighth of the long-term storage, for the same amount of work.

It's still a serious model

This isn't a stripped-down toy. DeepSeek trained it on 45 trillion tokens, taught it to read images, and says it stays competitive with the top models on the agent benchmarks that actually matter.

The honest caveats

One claim in DeepSeek's report — its biggest one — isn't proven by the data it shows. And DeepSeek itself is unusually clear about what could still break. This is a research report about a model you can't easily use yet, not a shipped product. The “8x cheaper” figure is DeepSeek's own, from its own tests.

What it means for you

Forget the benchmarks. The real story is that the cost of running an AI agent is falling fast — the same way compute and storage always have. As agents get cheaper to keep alive, handing them longer, bigger tasks stops being a luxury and starts being routine.

The bottom line

The race in AI is shifting from “who has the smartest model” to “who can run an agent the cheapest.” DeepSeek just made its move.