Since the launch of ChatGPT in November 2022, the AI industry witnessed a dramatic period of price deflation. Costs for accessing large language models dropped rapidly as competition intensified and infrastructure scaled. This trend led many to believe that AI would become increasingly affordable, even ubiquitous—so much so that Google CEO Sundar Pichai once predicted we’d soon have “intelligence, just like electricity,” implying it would be cheap and universally available. But that narrative is unraveling. The recent introduction of Anthropic’s “Fast mode” has shattered any illusion of declining prices. At $150 per million output tokens, this tier is more than ten times costlier than most mainstream alternatives. For context, many competitors charge between $1 and $10 per million tokens. This isn’t a minor price adjustment—it’s a record-breaking hike that puts high-performance AI out of reach for the average user, researcher, or small business. Dario Amodei, co-founder and CEO of Anthropic, famously expressed skepticism about AI’s ability to solve inequality and drive broad economic growth in his essay Machines of Loving Grace. Ironically, his company’s latest pricing move may be exacerbating the very problem he warned about. By setting a new benchmark for expensive AI access, Anthropic is reinforcing a trend where only well-funded corporations and wealthy individuals can afford cutting-edge models. So why aren’t prices falling? The answer lies not in the companies’ greed, but in the fundamental economics of frontier AI. As models grow more complex, they require exponentially more compute, energy, and specialized hardware—especially for real-time, high-throughput inference. Training and running state-of-the-art models now demands massive investments in GPUs and data centers. These costs are not disappearing; they’re shifting to the user side. Moreover, the race to build “superintelligent” systems has created a new class of premium services—fast, reliable, and highly accurate—that cater to enterprise clients willing to pay a premium. These services are not meant for mass adoption. They’re designed for high-stakes applications like autonomous systems, financial modeling, and advanced research—where performance outweighs cost. The real danger isn’t just the high price tag. It’s the widening divide. As AI becomes more expensive, access becomes more exclusive. Small developers, startups, academics, and innovators in the Global South are increasingly locked out. This risks creating a new digital elite, where only a few companies control both the technology and the ability to use it. The myth that AI is getting cheaper is not just misleading—it’s dangerous. It distracts from the urgent need for policy, infrastructure, and innovation that can democratize access. Without intervention, the future of AI won’t be one of shared intelligence, but of entrenched inequality. The cost of progress should not be the exclusion of the many for the benefit of the few.