Running models, and what they cost

One figure about cost made me reconsider the line "leave AI until next year".

What LLMOps is

Picture the cost of a long-distance telephone call a few decades ago: expensive enough that people weighed up every minute. Then the technology brought the price down, and today a video call across continents is near enough free. The cost of running large language models is travelling the same curve. LLMOps is the set of practices for putting these models into real operation: tracking quality, speed, and above all cost by volume of use.

According to Stanford HAI's AI Index 2025, the cost for a model at roughly GPT-3.5 level to answer fell from about $20 to about $0.07 per million tokens, more than 280 times cheaper in around eighteen months, from the end of 2022 to the end of 2024.

How this applies to running a business

I often hear the reasoning that "AI is still expensive, let us deal with it later". But when cost falls that fast, the argument weakens. What is genuinely expensive is no longer the tokens, it is the time a team spends not yet familiar with the work, not yet knowing which tasks to hand over, not yet having a process.

So the thing worth investing in early is not the most expensive plan, but letting the team start on small, low-risk tasks and build up experience. When prices fall further and the team is ready, the business can widen immediately instead of starting from zero.

If cost were no longer the obstacle, what is the first task you would want your team to practise handing to AI?