Last month I was scoping out a new AI feature for a clinic management product. When I ran the backend cost estimate, it came in at roughly one-tenth of what I'd calculated six months earlier. My first thought was "great, budget headroom." My second thought was more uncomfortable: "what exactly did we change?"
Nothing.
The numbers are striking. Since early 2024, frontier model inference costs have dropped 10-50x on average. For older model classes, the reduction reaches 900x. Something that was a real budget line item eighteen months ago is now basically noise.
Everyone knows this. Most product teams are still operating on 2024 instincts.
Here's what I mean: in 2024, the question "should we build this with AI?" came with a real cost calculation attached. Token counts. Monthly active users. Margin impact. That was the right question for that moment.
But that moment is over.
When inference cost drops, there are three things you can do with the savings:
1. Pocket it. Same features, lower infrastructure cost, higher margin. Easy. Temporary.
2. Build more. Ship everything that was previously too expensive. Sounds good, usually becomes feature bloat.
3. Pass it to customers. Lower your prices or unlock tiers for more users. This is the hard one.
In practice, almost no one picks option 3. The logic is "why lower price if they're already buying?" That logic holds in the short run. In the medium run, you're writing an invitation for competitors.
Concrete example. I work in health SaaS — software for small clinics. Six months ago, we made a decision to put certain AI-assisted features in the premium tier because the per-use cost was meaningful. That cost is now gone.
But those features are still in the premium tier.
Why? Because nobody schedules the meeting where you say "costs dropped, let's rethink the access layer." The product team is buried in the backlog. Pricing decisions feel like someone else's job. And the cost reduction just quietly flows into the P&L.
This isn't a backlog problem. It's a strategy disconnect.
Here's what I keep coming back to: when AI costs drop, the thing that actually needs to change is what I'd call the access layer of your product.
Access layer: which users can reach which value, under which conditions? If cost is no longer the constraint — and increasingly it isn't — what are you actually using to make that decision?
I've asked this to a lot of PMs. Most answer with "that's how the packaging works." Which tells me the packaging was set by an assumption that no longer holds.
The real PM question in 2026 isn't "how much does it cost to run this feature?"
It's: "now that this cost is gone, who should capture the value — us, our customers, or a competitor who figures it out first?"
That question belongs in strategy conversations, not pricing reviews. Because the answer will determine your competitive position over the next 18 months more than almost any feature you ship.
A few weeks ago I looked at the pricing page of a competing health software. Four features that were "premium" a year ago had moved into their base tier. Not because they found better technology. We're all hitting the same APIs. Because they made a decision: costs dropped, we're passing it on.
We haven't made that decision yet.
That's where I'm sitting with this. Cost collapse doesn't automatically update your strategy. Updating strategy is a separate muscle. And in most product organizations, it's the one that doesn't get exercised.