The AI pricing you budgeted against last year isn’t the AI pricing you’d get quoted today.
If your business passed on an AI project in the last year or two because the cost didn’t justify the outcome, that decision is worth revisiting. Not because AI is suddenly cheap across the board, but because the specific kind of work most business projects are made of has gotten meaningfully less expensive to run, and the tools for building it have matured alongside the price drop.
That’s easy to miss if the number in your head is still the one from your last conversation about AI budgeting. It’s worth understanding what’s actually changed before deciding whether last year’s “no” still holds.
What’s actually driving AI costs down?
Two things, and they’re related.
A Chinese AI lab called DeepSeek forced the issue by releasing a reasoning model that matched the leading US labs’ flagship performance at a fraction of the cost. That triggered direct price competition that hasn’t let up since. Major labs have kept responding with cheaper, faster options rather than settling back to where they started.
At the same time, the industry has shifted away from pricing AI as one flat rate and toward pricing it in tiers: cheap, fast tiers for routine, well-defined work, and pricier tiers reserved for tasks that need real multi-step reasoning. That shift didn’t happen in a single moment and isn’t finished. It’s an ongoing dynamic, which means the routine AI work behind most business use cases (classification, extraction, drafting, first-pass analysis) likely costs less to run today than it did the last time you priced it out, and may keep moving.
We dug into the economics on AI Unplugged Episode 15
Why doesn’t “AI got cheaper” tell the whole story?
Here’s the part worth being upfront about: at the frontier, AI hasn’t gotten universally cheaper. Every time a lab’s flagship model gets smarter, it tends to get more expensive, not less.
What’s mostly happened is a split. Routine tasks run on cheap, fast tiers priced a fraction of what a single flat rate used to cost. The hardest problems still run on premium, reasoning-heavy tiers priced at a real premium. That two-tier structure is the new default shape of the market, not a temporary discount.
Often, the tier just below the flagship model is where the movement is. OpenAI cut prices 50% on Sol and Luna, the two tiers under its flagship Astra model, while closing more of the performance gap to it. Anthropic’s Opus 5.5 launched around 20% cheaper than the Opus model it replaced, while closing in on the performance of Fable, Anthropic’s own flagship tier. In both cases, the flagship itself held its price. The tier below it didn’t, and it kept getting more capable while it dropped.
So the right question isn’t “did AI get cheaper,” it’s “which parts of my project actually need flagship pricing or higher reasoning,” because lower tiers keeps closing the gap for less.
How should you choose between the cheap tier and the reasoning tier?
There’s no clean formula yet. Which tasks genuinely need frontier-level reasoning and which can run on a cheap, fast model isn’t always obvious from the outside, and the honest answer usually comes down to testing it against your own use case rather than following a general rule.
That’s starting to change. The next generation of AI systems is being built with routing options. The system itself decides, per request, which tier a task actually needs based on its difficulty, rather than a single blanket setting. Until that’s standard everywhere, the practical move is to treat tier selection as something worth testing deliberately.
What’s worth re-pricing today?
If a project got shelved because the cost didn’t justify the outcome, run the math again. Specifically, not in the abstract. Most projects are made up mostly of routine work with a smaller slice of genuinely hard reasoning. The routine slice is where the price has moved the most, so a project that looked expensive across the board a year ago might look very different priced tier by tier today.
The businesses getting the most out of this shift aren’t assuming AI got cheap everywhere. They’re going back to last year’s “no” and finding out which parts of it can run on which tiers.
Ready to see what your shelved project would cost today?
Innovative Solutions helps organizations price AI work against current reality, not last year’s assumptions. Matching the right pricing tier to the actual difficulty of the task. Reach out to revisit a project that didn’t pencil out before.
