Two years ago, the owner of a small translation agency looked into using AI to produce first drafts of documents. The quality was patchy and the cost of running large volumes through the best models seemed high for a business her size. She decided to wait.
When she looked again, the picture had changed dramatically. Models of similar or better quality were available for a fraction of the price, some could run on an ordinary laptop, and her existing software had added AI features at no extra cost. Something that had looked like a luxury for large firms had become affordable for her.
One of the most important trends in AI is the rapid fall in what it costs to use capable models. For small businesses, that trend may matter more than any headline about the latest product launch, because it changes what’s affordable and practical to try.
One Like insight: When a technology gets dramatically cheaper, the question changes from “can we afford it?” to “what’s worth doing with it?”
The collapsing cost of AI
Each year, Stanford University’s Institute for Human-Centered Artificial Intelligence publishes the AI Index, one of the most comprehensive independent reports on the state of AI. Its 2025 edition highlighted a striking trend in costs.
The cost of using an AI system that performed at roughly the level of GPT-3.5, the model behind the original ChatGPT, on a standard benchmark fell by more than 280 times between November 2022 and October 2024. In other words, a level of capability that was cutting-edge and relatively expensive in late 2022 had become dramatically cheaper within two years.
The report also noted that smaller models had become far more capable, so that much smaller and cheaper systems could match the performance of earlier, much larger ones. And the gap between the best openly available models and closed commercial ones had narrowed considerably, giving businesses more choices.
These figures describe the cost of running AI models, mainly for developers and companies using them at scale, rather than the subscription prices individuals pay for consumer tools. Costs for the most advanced models can still be high, and the energy and infrastructure required to train and run AI at scale remain significant. But the overall direction is clear: capable AI is becoming much more affordable and widely available, including for businesses with modest budgets.
What cheaper AI means for small businesses
Falling costs change the practical calculation in several ways.
Experiments become cheap. Trying an AI tool on a real task, such as drafting product descriptions, summarising documents or answering common questions, now often costs little or nothing beyond time. That makes small, careful experiments sensible, as explored in Most Good Ideas Don’t Work. Test Before You Trust.
AI is arriving inside existing tools. Many accounting, email, design, booking and office software packages now include AI features. Before buying a separate tool, check what you already have.
Waiting has a cost too. If competitors use cheaper AI to respond faster, produce more content or serve customers better, standing still can mean falling behind, even if you avoid early mistakes.
The bottleneck shifts to judgement. When the technology is cheap, the scarce resources become clear thinking, good data, well-designed processes and people who know how to use and check AI. As explored in AI Makes Good Work Better, and Some Work Worse, knowing where AI helps matters more than access to it.
Old assumptions need revisiting. If you decided a year or two ago that AI was too expensive or not good enough for a task, it may be worth looking again.
[Infographic placement] AI got cheaper, fast. A simple downward line chart showing the relative cost of using a model at roughly GPT-3.5 level, falling from 100 (November 2022) to under 1 (October 2024): more than 280 times cheaper. Caption: Stanford HAI, AI Index Report 2025. Costs refer to running models, not consumer subscription prices.
Revisiting your AI decisions
Make a short list of tasks in your business where you’ve previously considered AI and decided against it, or never got round to trying. For each, spend half an hour testing what’s available now, starting with features in the software you already pay for. Note which tasks show genuine promise and which still don’t work well. Revisit the list every six months, because capabilities and costs are changing quickly.
A few questions can help:
- Which tasks did we rule out for AI because of cost or quality, and when?
- What AI features already exist in software we pay for?
- What small, low-cost experiment could we run this month?
- Do we have the skills and processes to use cheaper AI well?
Which brings us back to the translation agency and its owner’s decision to wait. Two years on, the cost barrier had largely gone. What mattered now was how she used AI: for fast first drafts, checked and refined by skilled translators who understood nuance and context. Cheaper AI didn’t replace her team’s expertise. It made that expertise go further, at a price a small business could comfortably afford.
Sources
- Stanford Institute for Human-Centered Artificial Intelligence (HAI), AI Index Report 2025.
