AI Costs Fall 13x a Year, Leaving Moore's Law Far Behind 2026

AI costs are plummeting at a rate that makes Moore's Law look sluggish, according to a new study from Epoch AI, a nonprofit research institute in San Francisco. The analysis finds that the cost of achieving a given level of AI performance has been falling by about 47% each quarter since 2023 — roughly 13 times cheaper every year.

That is a staggering pace, especially when compared to the decades-long rhythm of classical computing, where performance per dollar has historically doubled every 18 to 24 months. In other words, AI is not just improving; it is becoming radically more affordable at a speed the tech industry has never seen before.

What the Epoch AI Study Actually Found

Epoch AI's research tracks the cost required to reach a fixed level of AI capability over time. Rather than looking at the price of the latest flagship model, the study asks: how much does it cost today to perform a task that a model could handle a year ago? The answer, according to the data, is about 13 times less.

This methodology matters because it separates raw price tags from actual performance. A new top-tier model may cost more per token than its predecessor, but it also does far more. When you adjust for capability, the effective cost of AI is collapsing.

Why the Headline Number Is 47% per Quarter

The 47% quarterly decline is the engine behind the 13x annual figure. It means that every three months, the cost of reaching a specific performance threshold is nearly halved. Over a full year, those quarterly cuts compound into an order-of-magnitude improvement.

Epoch AI is careful to frame this as a trend observed since 2023, not a permanent law. Still, the consistency of the decline suggests structural forces at work — better model architectures, more efficient training, and smarter inference — rather than a one-off fluke.

Why AI Costs Are Falling So Fast

Several forces are driving the collapse in AI costs. The most important is that newer models are simply more efficient than their predecessors, delivering the same or better output with fewer computational resources.

• Architectural efficiency: Newer models achieve more with less compute, reducing the cost per task.
• Inference optimization: Techniques like quantization and distillation shrink models without sacrificing much performance.
• Hardware improvements: Specialized AI chips are getting faster and more power-efficient.
• Competitive pressure: Labs like OpenAI and Anthropic are racing to offer better performance at lower prices.
• Scale effects: Larger training runs produce models that generalize better, lowering the cost of specialized tasks.

Anthropic's own pricing illustrates the point. The company says its Opus 5.5 model matches the performance of Fable 5.1 on many tasks but costs less than Opus 5 — $4 per million input tokens and $20 per million output tokens. That kind of generational leap in price-performance is becoming routine.

The User Perception Gap

If AI is getting so much cheaper, why does it feel like your subscription credits vanish faster than ever? The answer lies in how users experience AI. As labs release more powerful models, they often consume more tokens per query or encourage heavier usage, so the sticker price of a plan may not reflect the underlying cost decline.

In practice, the Epoch AI finding is about the cost of a fixed capability, not the cost of your monthly bill. A task that required a costly frontier model a year ago can now be handled by a cheaper, more efficient model — but users may be opting for the newest, most capable (and most expensive) tier instead.

What This Means for Businesses

For companies building on AI, the implications are enormous. Tasks that were once too expensive to automate — like processing millions of documents or running real-time analysis — are becoming economically viable. The falling cost of AI is a tailwind for adoption across every industry.

It also changes competitive dynamics. Startups can now access capabilities that were previously the preserve of tech giants, leveling the playing field and accelerating innovation.

How AI's Cost Curve Compares to Moore's Law

Moore's Law, named after Intel co-founder Gordon Moore, described the doubling of transistor density roughly every two years. It became a shorthand for the steady, predictable pace of computing progress. AI's cost decline is far steeper.

The comparison is not perfect — Moore's Law tracked hardware, while Epoch AI tracks the cost of a capability — but the contrast is striking. AI is compressing decades of cost reduction into years.

What to Watch Next

The key question is whether this pace can continue. Epoch AI's finding is a trend, not a guarantee, and there are physical and economic limits to how cheap AI can become. Energy costs, chip supply, and data constraints could all slow the decline. The background to this story is set out in AI's Green Algorithm: How Software Cuts Energy Costs. A related thread runs through Antibiotic-resistant bacteria leaving children vulnerable: global. Earlier coverage in Ai Video: latest articles shows how the argument developed.

Another wildcard is regulation. As AI becomes more capable and more pervasive, governments may impose rules that affect pricing and access. How those policies evolve will shape whether the cost curve keeps bending downward. A closer look at the same subject appears in Gemini 4 Launch: Why Almost No One Can Use Google's New AI Model. Anyone following the wider picture can start with AI Torture Chamber: GitHub Project Sparks Ethical Debate on AI Pain. The same theme surfaces in AI Model Griffin Claims to Pass Video Turing Test, Nvidia Disagrees, which takes a different angle.

For now, the message is clear: AI is getting cheaper at a rate that defies historical precedent. Businesses and individuals who plan for a world of abundant, low-cost AI will be better positioned than those who assume today's prices are permanent. Context on how this played out before sits in California Bans Child Marriage, 32 States Still Allow. A companion piece, AI Costs Fall 13x a Year, Leaving Moore's Law Far Behind, approaches the same subject from another angle.