DeepSeek Flash keeps leapfrogging its own Pro tier

DeepSeek announced v4.1 Flash via a banner on their usage dashboard, claiming it's cheaper and more capable than v4 Pro. They're also routing all Pro requests to Flash before the official launch, which is an unusual move that signals high confidence in the new model. Multiple HN commenters said they've already switched almost entirely to Chinese flash models for daily work, citing pennies-per-answer costs and reliability on full-stack projects.

The pattern across several threads today, including the local model benchmarking discussion on Qwen3.8 27B quantizations, is that capable AI is getting aggressively cheap. The Qwen thread found that 4-bit quantization holds up well, meaning you can run strong models on consumer hardware. DeepSeek's move reinforces this: you can now get near-frontier quality for almost nothing if you're willing to use a Chinese-hosted model.

The risk people are dancing around but not naming directly is dependency. One commenter compared it to Hansel and Gretel following a trail of cheap tokens, with DeepSeek as the witch. The open-weights angle matters here: people are hoping v4.1 Flash is open weights so it can run locally, reducing that dependency.