AI Costs Are Falling. Why Your AI Bill Keeps Rising

Anthropic released Claude Opus 5 in July at exactly the same list price as the model it replaced. Five dollars per million tokens going in, twenty five coming out. Unchanged.

The interesting number is underneath the sticker. On CursorBench, a coding benchmark, Opus 5 landed within half a percent of the top-scoring model while costing roughly half as much per task completed.

Same price on paper. Half the work to get the same result.

This is not an Anthropic story. Forbes ran a piece on July 28 about token costs falling so fast that enterprise AI vendors are getting squeezed on their own margins. The capability you were quoted for last year now costs a fraction of that.

Which should mean your AI spending is going down.

Go look at your invoices. It probably isn't.

I want to slow down here, because the reason is not what most people assume. Nobody is playing a pricing trick on you. The math is more ordinary than that, and more interesting.

When something gets cheap, you use more of it.

Cheaper tokens make more things economically possible. So a workflow that ran once a week now runs on every incoming email. An assistant that answered questions now sits open all day. Agents retry when they fail, and retries cost money. Longer context windows mean bigger prompts, and bigger prompts mean bigger bills.

The unit price falls. The unit count explodes. Your total goes up.

That's the mechanic. The consequence is the part worth your attention.

For two years, cost was doing a job for you that you probably didn't notice. It was filtering. When someone suggested pointing AI at your entire customer email history, the price tag ended the conversation. Cost was a bad filter, but it was a filter.

That filter is dissolving, and most businesses have not installed anything in its place.

So here are two things worth doing, and both are small enough to finish this week.

First, your "too expensive" list is out of date. Whatever you priced out six months ago and walked away from should be re-quoted now. Transcribing every sales call. Summarizing a decade of project files. Drafting a first pass on every proposal. Pull up the one you most regretted saying no to and get a current number on it. Prices in this market move faster than your assumptions do.

Second, find out what your most-used AI workflow costs per run. Not per month. Per run. Most owners can tell you their monthly AI total and cannot tell you what a single invoice summary or a single customer reply costs them. That per-run number is the one that tells you whether growth is going to be fine or ugly.

Once you have it, multiply it by ten. Then by a hundred. If either number frightens you, you have found something to put a limit on before it finds you.

There's a broader version of this that goes beyond spending.

When a thing gets cheap, we stop asking whether it's worth doing. Film made people think before the shutter. Nobody thinks before the shutter anymore, and most of us have four hundred photos of the same lunch.

That's harmless with photos. It's less harmless with the reports nobody reads, the automated follow up emails your customers have started ignoring, and the meeting summaries that pile up because generating them costs almost nothing now.

The trap isn't overspending. It's producing more of something without getting any better at deciding what's worth producing.

Cheap capability doesn't make any of those decisions for you. It just removes the excuse you were using to avoid making them.

So the question I'd sit with this week isn't what your AI stack costs. It's which parts of your business would actually get worse if they got faster.