DeepSeek-V4.1-Flash outruns V4-Pro, reroutes its API calls

DeepSeek's pitch here is unusual: the smallest member of a new architecture family is being positioned above the company's own flagship. V4.1-Flash is a mixture-of-experts model with 552 billion parameters, close to double the 284 billion in V4-Flash, but a new causal encoder-decoder design keeps only 8 billion parameters active while reading a prompt and 16 billion while generating output. Much of the engineering went into shrinking the key-value cache, which the technical report says is stored in a four-bit floating-point format at 890 bytes per token, about a quarter of what V4-Flash needs, with persistent SSD cache storage down to roughly an eighth of the previous generation.

The pricing move is the practical part. Rather than ask developers to switch, DeepSeek is switching them: from Sept. 14, V4-Pro API requests are answered by V4.1-Flash and billed at the smaller model's rates until a V4.1-Pro version launches, and the retired V4-Flash and August's experimental vision model already route there. Image understanding, previously only in that experimental release, is now built into the model itself.

The launch lands the same day Anthropic named DeepSeek in its threat intelligence report as one of seven China-based labs it says ran distillation campaigns against Claude, attributing more than 12.1 million exchanges over 14 days in July to DeepSeek. DeepSeek's ability to hold developers on the rerouted API may hinge on whether the promised V4.1-Pro arrives soon, and on whether buyers weigh Anthropic's claims alongside the roughly 70% output-price cut.