AI Just Got Cheaper. Here Is What to Do

The price of running an AI model has dropped hard. By one widely cited measure from Stanford's 2025 AI Index, the cost of a query at the level of GPT-3.5 fell from about twenty dollars per million tokens in late 2022 to roughly seven cents by late 2024. That is not a discount. That is a different category of cost.

At Bace Agency, we have watched that shift change what is actually possible. Automation that was too expensive to justify for a North Shore agency two years ago is now cheap enough to build and run every day. Heavy document processing, always-on intake, a review of every file before it moves: work that used to price itself out now pays for itself.

So when an owner asks whether to wait for prices to fall further, my answer is no. The model price was never the part that cost you anything.

Why Prices Keep Falling

This is not a sale that ends Friday. Prices are falling because the hardware gets cheaper every year, the models get smaller for the same quality, and the vendors are in a fight for your business. Anthropic, for example, has held or cut prices on its Claude Opus and Sonnet lines across several flagship launches, including a steep cut to the Opus tier at the 4.5 release. That is unusual. Most products raise prices on a new top model.

I wrote a fuller explanation of the mechanics in why AI keeps getting cheaper. The short version: this trend is structural, not promotional. So waiting does not catch a bottom. There is no bottom to catch. There is only later.

Why Waiting Does Not Pay

Here is the part owners miss. On a real project, the model bill is the smallest line item. I broke down the full picture in the real cost of an AI project, but it comes down to three things, and none of them are the API.

First, figuring out the workflow. Sitting with your team, watching how intake actually happens, where the renewal list really lives, who re-keys the same client data into three systems. Second, building it and connecting it to the tools you already run. Third, getting people to trust it and use it Monday morning. That last one is the hardest, and a cheaper model does nothing for it.

None of that gets cheaper because tokens got cheaper. An insurance agency that waits a year to automate its intake does not get a discount on the project. It gets a year of staff re-keying data by hand, then starts the same build twelve months behind. One North Shore P&C agency we worked with cut about 22 hours a week of data entry, and the error rate fell from 6.1% to 0.3%. The value was the hours, not the token price.

What to Do With Cheaper AI Now

Cheap tokens do change one real thing: tasks that were too expensive to run on AI a year ago are worth running today.

When a model cost twenty dollars per million tokens, you saved it for high-value work. At seven cents, you can point it at the boring, high-volume tasks you used to skip. Read every inbound email and sort it. Summarize every long PDF before a partner opens it. Draft the first version of every renewal notice. Check every invoice against last month's. The math that said "not worth it" now says "run it."

So do two things. Turn AI loose on the cheap, repetitive tasks where being wrong is low-stakes and a human still reviews the output. And separately, pick one real workflow that costs your firm hours every week, and build that one properly. The first move is almost free to try. The second is where the actual return is.

How to Get Started

What This Does Not Replace

One honest limit. On anything that touches a client, a contract, or money, a person still reviews the output before it goes out. We build that review step in from day one. It is what makes the system safe to run at volume.

Past that, the point is simple. The model is cheap. The architecture is where the value lives. You are not buying tokens. You are buying back the hours your staff loses every week. In two of our North Shore engagements that meant about 22 hours a week reclaimed for a P&C agency and about 32 hours a week for a family office (case studies). That is what we build at Bace Agency.

So the real question is not whether AI will get cheaper. It will. The question is how many hours your team loses every week that you wait. The firms that move first this year will spend the next one running leaner than the ones still reading reports. Instead of reading another report, look at your own numbers with me. Book a 30-minute operational review. I will personally walk through your workflows and tell you exactly which tasks can come off your plate first. In person on the North Shore or by video. Start here, or take the AI Readiness Quiz first.

Frequently Asked Questions