You’ve seen the chart. DeepSeek V4 Flash 0731 matches GPT-5.6 Luna on ARC-AGI — the hardest AI reasoning benchmark — at 1/20th the cost. It feels like a steal. A bargain. A market correction.
It’s none of those things.
Here’s what actually happened: the floor just fell out of the AI pricing model, and almost nobody is looking at the right signal.
The dollar cost is a distraction. The real story is the efficiency race that will upend the entire AI industry.
Let me show you the weirdest part. DeepSeek’s pricing is so counterintuitive that max reasoning is cheaper than high reasoning. Think about that. You pay less to get the model to think harder. That’s not a typo. That’s a sign that the old rules — more intelligence = more compute = more money — are dead.
One commenter nailed it: ‘Kimi K3 was an interesting model only a month ago, and now we’re looking at the same performance for 1/20th of the price. Wild how fast this is advancing.’
But here’s the twist. That price tag? It’s contaminated. VC subsidies, economies of scale, and aggressive inference optimizations are all baked in. The real number nobody is talking about is the underlying efficiency — the FLOPs per reasoning step, the training data leverage, the architecture improvements that make the same capability cost a fraction of what it did last month.
Capability is becoming a commodity. Efficiency is the new moat.
This isn’t a pricing war. It’s a fundamental shift in what ‘winning’ means in AI. The companies that build the leanest, most efficient models — not the flashiest demos — will own the next decade. The ones still bragging about benchmark scores at any cost are already behind.
So next time you see a chart comparing prices, stop. Ask yourself: what’s the real cost per reasoning step? What’s the forward-pass FLOP count? Because the price you see today is a snapshot of a market that’s moving faster than our models can price.
And that’s both exhilarating and terrifying.
FAQ
Q: Is this just a pricing war?
A: No. The price is distorted by subsidies and scale. The real signal is the underlying efficiency gains — FLOPs, training data, architecture. A pricing war is temporary; the efficiency race is structural.
Q: What does this mean for my AI product decisions?
A: Stop optimizing for today's price. Start optimizing for cost per reasoning step. The model that costs $1 today might cost $0.05 next month. Build your stack to be model-agnostic and efficiency-first.
Q: Isn't the price drop just a subsidy bubble?
A: Partially. Subsidies inflate the apparent cheapness. But the underlying efficiency improvements are real and accelerating. When the subsidies fade, the models with the best FLOPs efficiency will survive — and the rest will vanish.