AI Pulse · Business Models & Winners

The price of using large language models is dropping quickly. Silicon Data’s token-price index fell from a peak near $2.07 per million tokens in May to $0.97 by August 31. That’s close to a 60% decline in a few months.

Three forces are behind it. Cheaper open-weight models are giving companies more options beyond the biggest providers. Auto-routing tools can send each task to the lowest-cost suitable model. And more firms are moving inference, meaning the actual running of AI models, onto hardware they own.

For a business owner using AI internally, lower token prices should reduce the cost of customer support bots, document processing, sales research, internal search, and other high-volume workflows. It also means you should revisit any AI cost assumptions you made earlier this year. A use case that looked too expensive in May may now be workable.

The harder hit will be businesses whose main product is simply reselling model access and charging by the token. If their input costs keep falling, customers will expect lower prices too. Unless those firms offer something customers can’t easily reproduce, such as a specialised workflow, proprietary data, strong integration, or a trusted service layer, their margins come under direct pressure.

This is the kind of pricing shift we build into an AI command centre, where model choice and routing should be managed as an operating decision, not left to chance.