By Jon Michel Greenwood, Interim CPTO, Chameleon Collective
Somewhere in most companies sits a shelved AI idea. A year or so ago, someone priced out automating support triage, invoice matching, or contract review. The model costs sank the AI business case, and the deck went into a folder marked "revisit later."
Later has arrived sooner than anyone planned. AI prices are falling so fast that last year's "too expensive" may be this year's competitive advantage.
Since 2023, the cost of reaching a fixed level of AI benchmark performance has dropped about 47% per quarter, or roughly 13x per year. The figure comes from Epoch AI's September 2026 report, The Plunging Price of Thought, by Luke Emberson and David Roodman.
Epoch set that pace against other transformative technologies, each measured over its own era of falling prices:
Epoch's own analogy: it is like a $50,000 new car falling to $69. The drop is steepest right after a capability first appears, about 66% per quarter for newly state-of-the-art performance, easing to about 32% per quarter two years on. Researchers suggest this may be the fastest decline of any transformative technology, and earlier, rougher evidence points to a similar curve back to GPT-3's commercial launch.
That curve deserves attention, and discipline in how we read it. These are prices for comparable benchmark performance. An entire AI operating budget will not fall at that rate.
Benchmarks are a stand-in for useful work, not the work itself. Epoch's method also assumes a user who always switches to the cheapest capable model, and real teams rarely move that fast.
A falling API price also hides its cause. It could be engineering efficiency, competitive pressure, or a provider accepting losses to win customers. Subsidized pricing and sustainable production economics look identical on an invoice.
Some pressure runs the other way. Memory prices face continued upward pressure as AI server demand strains supply. Hardware constraints can coexist with improving AI price-performance.
My own bias leans toward revisiting AI decisions often. Still, staying put can win. Swapping a model embedded in a regulated, validated workflow may trigger costly re-certification, and that cost can erase the savings.
As an Interim CPTO, my takeaway is simple: an AI business case should carry an expiration date, just like the pricing it was built on.
Reopen the "no" pile. A workflow that failed the economics test six months ago deserves another look.
Retest the incumbent model. A model chosen last year should be tested against newer, cheaper alternatives on your own tasks, not just public benchmarks.
Set a review cadence. On the average curve, a cost estimate two quarters old may overstate today's price for the same performance by more than three times. Quarterly reviews keep decisions in step with the market.
A token price is an ingredient, not the result. The number that matters is cost per successfully completed task, with everything around the model included:
Cost per completed task = (model + integration + human review + retries and errors) / tasks completed successfully
A cheap answer that requires expensive correction is still expensive. A slightly pricier model that gets it right the first time can win on this measure.
Lower unit prices make more applications viable. That can expand total AI spending even as each individual task gets cheaper. Plan the budget around the new use cases cheaper AI will unlock, not just the savings on today's workload.
The opportunity is substantial. Products, services, and customer experiences that could not justify their cost may now become commercially viable.
But if model access keeps getting cheaper for everyone, cheap AI alone will not set anyone apart. The advantage shifts to what competitors cannot buy off a price list: proprietary data, AI woven into real workflows, trusted customer relationships, and an organization willing to revisit its assumptions as quickly as the technology changes.
Two questions worth putting to your leadership team:
Which "too expensive to automate" decision would you revisit if comparable AI capability became ten times cheaper? At the current pace, that is less than a year away.
If model access keeps getting cheaper for everyone, what will make your product or business meaningfully harder to compete with?
Most companies do not need a new AI strategy every quarter. They need the habit of checking whether last quarter's answer still holds.
That is the work Chameleon Collective's fractional technology and product leaders do every day: pressure-testing AI business cases, comparing models on real tasks, and building the operating rhythm to keep pace with falling AI costs.
If there is a "too expensive" AI project sitting in your folder, it may be worth a fresh look. Happy to compare notes anytime; connect with me on LinkedIn or through my Chameleon Collective profile.
Epoch AI, The plunging price of thought, Luke Emberson and David Roodman, September 22, 2026
Epoch AI: The Cost of AI Performance Falls 13x Every Year, FourWeekMBA
Epoch AI says AI prices fall 47% a quarter at the same performance level
Newsletter
Perspectives on transformation, leadership, and growth from 280+ senior operators. No fluff, no spam.
