AI Gets Twice as Capable Every 12 Months and Half as Expensive Every 18. Here's What That Actually Changes.

Something unusual is happening with AI pricing that almost nobody is discussing clearly: the cost per unit of AI capability is falling faster than almost any technology in history. When GPT-4 launched in March 2023, OpenAI charged $30 per million input tokens. By March 2026, GPT-5.4 — a model that outperforms GPT-4 across most benchmarks — costs $10 per million tokens. That is a 3x price reduction alongside a capability increase. Gemini 3.1 Flash-Lite, released in March 2026, costs $0.25 per million tokens. At the consumer subscription level, the same $20/month you paid for ChatGPT Plus in 2023 now buys access to a model that outperforms what the best AI labs could produce three years ago.

The Historical Comparison

The closest historical analog is semiconductor pricing — Moore's Law, which described the doubling of transistor density roughly every two years, eventually making computation essentially free for most purposes. AI capability is following a similar curve, but faster. The specific drivers are hardware improvements (each new GPU generation delivers significantly more AI compute per dollar), software efficiency gains (better training methods require fewer compute cycles for equivalent output), and competition among providers racing to offer the best price-performance ratio. The result: AI capability that cost $100 to purchase in 2023 costs approximately $10 to purchase in 2026, and will likely cost $2-3 to purchase by 2028.

What This Means for Local AI in 2026

One of the most significant developments is Intel's Arc Pro B70 GPU: 32GB of VRAM for $949. VRAM is the primary constraint for running AI models locally. In 2023, running a capable local AI required expensive enterprise hardware. In 2026, a $949 GPU lets you run quantized versions of models like Qwen 3.5 27B at 4-bit precision — models that would have required a $10,000+ workstation two years ago. Local AI means no subscription fees, no rate limits, no data sent to external servers, no advertising, complete control over your AI environment. The economics of local AI are crossing a threshold where mainstream adoption becomes realistic for professionals and serious enthusiasts.

What This Means for Your Career Decisions

• Skills built on AI capability have a faster expiration date than skills built on AI judgment. Knowing how to use GPT-4 was a valuable skill in 2023. By 2026, the capability has improved so much that the specific skill of prompting that model is largely obsolete. Skills that age better: understanding when to use AI versus when not to, how to evaluate AI output quality, and how to integrate AI into complex workflows requiring domain expertise.
• The tasks that are currently expensive to automate will become cheap to automate automatically. Legal document review, financial modeling, code review — these are currently expensive because AI inference costs money. As inference costs fall 3-5x every two years, the economic case for automating these tasks strengthens without any additional development work required.
• Individuals who learn to orchestrate multiple AI models will have an asymmetric advantage. When AI capability costs essentially nothing and is available to everyone, competitive advantage comes from how intelligently you deploy it. The professional who knows which model to use for which task, how to chain outputs between models, and how to evaluate quality will consistently outperform those who use one tool for everything.
• Local AI changes the privacy equation for sensitive work. As local AI becomes accessible, professionals handling sensitive data — lawyers, therapists, financial advisors, healthcare providers — have a viable path to AI-enhanced productivity without cloud-based data security concerns.

The Most Important Implication

The falling cost of AI capability means that within 3-5 years, the distinction between AI-assisted work and regular work will largely dissolve. Just as the question 'do you use the internet for your work?' became meaningless by 2010 because the answer was universally yes, the question 'do you use AI for your work?' will similarly become moot. The cost curve is eliminating the access barrier. What remains is the judgment barrier — knowing how to use AI well, when to use it, and how to evaluate what it produces. That is a human skill that AI cannot yet replicate, and it is the skill worth investing in now.