The Cheap Think Machine: How AI Got Commoditized Faster Than Anyone Expected
Site Owner
Published on 2026-05-21
In 28 months, AI inference costs fell 600x. Here's why it happened so fast, what it actually means, and which layers will commoditize next.
The Cheap Think Machine: How AI Got Commoditized Faster Than Anyone Expected
In January 2023, GPT-4 cost about $60 per million tokens. By May 2025, you could run a capable reasoning model for under $0.10 per million. In roughly 28 months, the price of machine intelligence fell 600x. No industry analyst predicted that pace. Not even close.
This is the story of the most aggressive commoditization event in technology history — and why it caught nearly everyone off guard.
The Benchmark That Started a War
It began, as so many things do, with a leaderboard. When o1-preview landed in September 2024, it changed what "thinking" meant for AI. For the first time, a model didn't just retrieve — it reasoned. It spent tokens as a budget for deliberation, and on hard problems, it showed. MMLU, GPQA, coding competitions: the numbers looked like a step change, not a marginal gain.
The reaction was instant and predictable: every lab scrambled to build a "reasoning" model. But what came next was not a race to match OpenAI's quality. It was a race to match it at a fraction of the cost.
DeepSeek R1 fired the starting gun. Released in January 2025 with a fully open weights license and inference code, it posted reasoning benchmarks competitive with o1 at roughly 1/30th the API price. The community did what communities do: someone quantized it, someone else distilled it, a third person put it on a consumer GPU and ran it locally. The reasoning premium evaporated in weeks.
Why the Cost Curve Moved So Fast
Three forces collided simultaneously.