Decision guide
AI gets markedly cheaper every year, and that is measured, not guessed. So is waiting the cheaper move? Here is an honest framework: what the price collapse really means, what it does not solve, and what to base a decision on that still holds in a year.
In short: Yes, AI gets markedly cheaper every year. Waiting is still rarely the right call. A cheaper model always arrives. The real bottleneck is not the price, it is the time your organisation needs to learn where AI actually helps. Start with a small, real use case. And build on what stays, not on a model that will be dated within a year.
It is a fair question. If a model will do more for less in twelve months, why start today? We build Pinchy, an open-source governance layer for AI agents, so we are not neutral here. We still wrote this framework to be useful even if your answer ends up being "not yet". Because the numbers say something more interesting than the question suggests.
This is not a forecast. It is a measured trend. The most recent solid measurement comes from MIT: between April 2024 and November 2025, the price for the same benchmark performance fell to between a fifth and a tenth each year. What matters is what they measure: the price per completed task, not per token. The heavier token use of reasoning models is therefore already priced in.
Over a longer window the curve is steeper still. a16z calculated in late 2024: GPT-3-level capability cost about $60 per million tokens in late 2021. By late 2024 an open model delivered the same performance for 6 cents. That is a factor of 1,000 in three years.
These numbers compare equal performance, measured on the same tasks. This is not about ever-better frontier models, it is about the price of a fixed level.
Two technical forces drive this. The first is hardware, and right now it is the exception. Over two decades, compute per dollar fell by roughly 30% a year (Epoch AI, as of 2024). The memory shortage is currently interrupting that trend: DRAM contract prices rose 93% to 98% quarter over quarter in Q1 2026 (TrendForce). Buy hardware today and you pay more than last year.
The second force keeps going regardless. For the same performance, models now need only about a third of the compute each year. That is a property of the software, and it holds no matter what a GPU costs. The number is unusually solid: Epoch AI measured it in 2024, MIT in 2025 with a different method. Both land in roughly the same place.
Then there is the market. The moment an open model does for almost nothing what was expensive yesterday, no vendor holds its price. Intelligence becomes a commodity, and commodities get cheaper under competition.
Here is the flaw. There is always a better and cheaper model six months out. And again twelve months out. The curve does not flatten. If "better and cheaper" is your reason to wait, you wait forever, because something better is always coming.
And while you wait, the real thing happens elsewhere. The teams that started are learning. They can see where AI pays off in their own operation and where it does not, and which processes are worth it. They watch how their people work with these tools without trusting them blindly. You cannot download that head start. It only builds over time.
The expensive bottleneck is not the token price. What decides it is the time until your organisation genuinely knows how to use this. That time does not appear on its own. You have to take it.
The token price is falling. True. Your AI project does not automatically get cheap, because compute was never the expensive part.
What gets cheaper is what a machine does alone: computing, generating text, spotting patterns. What stays expensive is anything that needs judgement. Your data is messy, and someone has to sort it. Someone has to decide what an agent may do. Someone has to check that only what should go out actually goes out. And when something goes wrong, someone is liable for it.
That does not cost tokens, it costs work and responsibility. And that does not halve every few months. So the price collapse does not lower your project costs. It shifts them, away from computing and towards trust.
One part of your project loses value faster than any other: the model. The cost of a fixed capability halves roughly every few months. The model you pick today will be mediocre in a year.
Teams still build entire projects around that one model. The whole architecture hangs off it. Then the next one arrives, better and cheaper, and the rebuild starts over.
The model is not your foundation, it is the ingredient you swap out fastest. What stays is the layer around it. Who is allowed to trigger what? Which action lands in a log that cannot be altered afterwards? How is an agent stopped before it does damage? That layer does not age. It also carries every better model that arrives next year. Build your head start where it holds, and not on the part that halves every few months.
The decision comes down to five questions, not to price:
| Start now | Waiting is defensible | |
|---|---|---|
| Use case | concrete and small, already holds up today | just an idea for later |
| Where the value sits | in the process and the learning | in the newest model alone |
| Missing capability | none, today's models are enough | strictly required, does not work reliably yet |
| Capacity | a first project can be supported | none at the moment |
| Maturity | grows while the stakes are small | meant to appear at some point |
The difference is rarely the price. It is whether you have a real use case and the time to learn from it.
Prices can rise in the short term, for instance when compute capacity gets scarce. Over several years, competition beats any shortage. As long as open models keep arriving and efficiency keeps improving, the curve points down. A bubble changes the pace, not the direction.
The frontier model, yes, and that is measured too. The same MIT work tracks the opposite direction: the cost of running frontier models rises to between 3x and 18x per year, because models grow larger and spend longer on reasoning. Both hold at once: the price for the same performance falls, the price of staying at the frontier rises. You pay the frontier price only if you genuinely need the frontier. For most applications yesterday's model is enough, and it costs almost nothing.
Yes, and it is the most important caveat on this page. SWE-bench Verified poses tasks from real software projects. There, MIT measures roughly constant cost per run. The higher compute load eats the falling token price: models simply spend far longer per task. The more agentic the task, the weaker the effect. If you are building agents for long, open-ended work, do not count on falling prices to solve the problem for you.
Now for our own product. Pinchy is the layer that stays when the model underneath changes. It is an open-source governance layer and runs on OpenClaw. It handles per-user identity, grants permissions via allow-lists, and writes a signed audit trail. It is deliberately not tied to a model. When the next model is better and cheaper, you simply swap it. Your rules, permissions and logs stay put. That way you build something today that benefits from the price collapse instead of being overtaken by it.
FAQ
Yes, and that is measured, not estimated. The most recent measurement comes from MIT: between April 2024 and November 2025, the price for the same benchmark performance fell to between a fifth and a tenth each year. They measure price per completed task, not per token, so the heavier token use of reasoning models is already priced in. Two technical forces drive it: cheaper hardware and more efficient models. Then there is competition from open models.
Usually no. A better and cheaper model always arrives. If that is your reason to wait, you wait forever. The real bottleneck is not the token price. It is the time until your organisation has learned where AI holds up and how not to trust it blindly. Waiting is only defensible if you have no real use case today.
Not automatically. The price collapse affects compute, and compute was never the expensive part. What stays expensive is integration, data quality, judgement and governance. So the price collapse does not lower your project costs. It shifts them away from computing and towards trust.
That is exactly why. The model ages fastest, its cost halves roughly every few months. The layer around it stays: identity, permissions and a tamper-evident audit trail. That layer carries every new model that arrives, which is why it is the part that holds your investment.
It varies by task. MIT measures that between April 2024 and November 2025 the price for the same benchmark performance fell to between a fifth and a tenth each year. Over a longer window the curve is steeper: a16z calculated in late 2024 that GPT-3-level capability fell from $60 in late 2021 to 6 cents in late 2024 per million tokens, a factor of 1,000 in three years. On strongly agentic tasks, however, the effect barely lands.
Ongoing maintenance and governance, not tokens. Compute gets cheaper every year. What stays are integration, data upkeep, permissions and a verifiable audit trail. That part costs work and responsibility, and it does not halve every few months.
Pinchy is the open-source governance layer for OpenClaw: identity, allow-list permissions and a signed audit trail. Model-independent, self-hosted, free to run.
Or email us: info@heypinchy.com