A decisive truth, applied directly to the forehead of Western AI pricing models: Chinese AI lab DeepSeek just released V4, and it costs approximately the same as the lint you find between your couch cushions.
DeepSeek V4 Pro packs 1.6 trillion parameters — that’s trillion with a T, the same letter that starts “terrifying” — with 49 billion active at any given time. The Flash variant runs 284 billion parameters with only 13 billion active. Both come with million-token context windows, because apparently reading your entire life story in one gulp is now a baseline feature.
🤚 The Price Slap
Here’s where it gets beautiful. Input tokens: \$0.14 per million. For reference, that’s cheaper than a single gumball from one of those machines at the grocery store. OpenAI, Google, and Anthropic are presumably staring at their own pricing pages with the same expression you make when you realize you’ve been overpaying for car insurance for six years.
On reasoning benchmarks, V4 is approaching GPT-5.2 and Gemini 3.0 Pro territory. Not quite there, but close enough to make you wonder if the \$200/month subscription model has a shelf life shorter than grocery store sushi.
👐 The Mixture-of-Experts Situation
Both models use mixture-of-experts architecture, which — and we cannot stress this enough — is just a fancy way of saying “we have a team of 1.6 trillion specialists, but we only wake up 49 billion of them for any given question.” It’s like having a hospital with every possible specialist, but only paging the ones who are relevant. Efficient. Slightly unsettling.
🌿 The Gentle Awakening
The real story isn’t that DeepSeek is cheap. It’s that “cheap and nearly as good” is an extinction-level event for anyone whose business model is “expensive and marginally better.” The AI pricing wars have entered their Walmart phase, and some very well-funded companies are about to learn what competition feels like.
DeepSeek’s model was unavailable for comment, as it was busy processing your entire email history for the cost of a single pistachio.