On September 10, Chinese AI lab DeepSeek released V4.1-Flash with pricing that makes frontier models look like a luxury tax. Off-peak, it costs $0.15 per million input tokens and $0.60 output—half what DeepSeek charged last month and roughly 30 times cheaper than OpenAI's $10 input rate for GPT-6 Astra.
The performance held
V4.1-Flash is not a budget model. It outscores Claude Opus 5 and GPT-5.6 Sol on agentic and coding benchmarks, runs native vision, and ships with MIT-licensed open weights. On September 14, DeepSeek retired its V4 Pro model entirely, rerouting all API traffic to V4.1-Flash at the new lower price.
Why this is not good news
Cheaper is better—until it's not. When one provider undercuts the entire market by an order of magnitude, the only moat left is switching cost. Your evaluation pipeline now assumes DeepSeek. Your cost model breaks if you move. Your training data lives in Shenzhen.
OpenAI and Anthropic still charge $10-$50 per million tokens, not because their models cost that much to run, but because they can. DeepSeek just made that untenable. The next move is price war or lock-in, and both end the same way: you own less of the stack than you did yesterday.