Thirty Cents a Million: DeepSeek V4.1 Flash and the Collapsing Cost of Intelligence

The Fortnight That Broke My Price Assumptions

September 2026 has been a strange month for AI news, even by 2026 standards. In the first six days alone, all four major labs shipped a flagship: Claude Fable 5.1 and Mythos 5.1 from Anthropic — the poetry family got a 5.1 faster than most indie games get a patch — then Meta's Muse Spark 1.3, Google's Gemini 3.8 Flash, and OpenAI's GPT-6 Astra, whose staged rollout went badly enough that Sam Altman had to publicly apologize for it (Startup Fortune). CNBC gave the phenomenon a name, "model fatigue," and honestly, I felt it. Four keynotes in six days. My brain stopped downloading release notes around Wednesday.

And then, quietly, DeepSeek did the thing I keep forgetting to expect from them.

On September 10 they shipped V4.1 Flash, and the tagline reads like a dare: 更强、更快、更普惠 — stronger, faster, more affordable for everyone (DeepSeek). Here's the part that made me sit up straight: V4.1 Flash is a 552-billion-parameter MoE model with a brand-new Causal-Encoder-Decoder architecture, native vision understanding, open weights under an MIT license — and it is, by their own telling, the smallest model in the new architecture family. The smallest one. And it beats their own flagship V4 Pro comprehensively — on performance, on cost, on speed, on total runtime, per tests by multiple parties that the announcement page cites — which is why V4 Pro is being retired: from 04:00 UTC on September 14, every request to deepseek-v4-pro gets silently routed to V4.1 Flash and billed at the cheaper rate (changelog).

Let that sink in for a second, because I had to reread it too. The flagship didn't get outclassed by a rival lab this time. It got outclassed by its own little sibling, and then it got walked off the stage. (・∀・)

The fastest way I know to kill a flagship is to build something better and price it lower. Somebody should tell the big labs.