AI costs fell 47% a quarter. Your AI bill may still rise

On Tuesday night the research group @EpochAIResearch posted a chart with a big claim on top: "AI is getting cheaper more quickly than any other transformative tech in history." By their count, the cost of reaching a fixed level of AI performance has dropped about 47% per quarter since 2023. Over a year, that's roughly 13 times cheaper.

Epoch also compared it with earlier technologies. The decline runs four times faster than DNA sequencing costs fell, six times faster than compute, 18 times faster than lithium batteries, and 54 times faster than electricity in the century up to 1973. The post has passed a million views. For anyone deciding whether AI is worth paying for, it's a chart worth a look.

It doesn't mean your AI bill will shrink, though, and that gap is the part a business owner should pay attention to this week.

What Epoch actually measured

The full report, The plunging price of thought, follows five benchmarks: competition math, the very hard FrontierMath set, the GPQA Diamond science exam, chess puzzles, and mystery game puzzles. Cost here means the real cost of running each task, so the price per token times the number of tokens the model burns.

The science exam gives the cleanest example. A 75% score on GPQA Diamond cost about 30 cents a question with OpenAI's o3. GPT-5.6 Luna now gets there for four hundredths of a cent, a 725-fold drop in about 18 months.

Prices fall fastest right after a new level of performance first shows up, at about 66% per quarter. Two years on, the decline slows to about 32% per quarter, which is still quick by any normal standard.

The caveats Epoch flags itself

The authors are upfront about the limits. Labs may train specifically for popular benchmarks, so scores can climb faster than real-world usefulness. There are only three years of data. And they admit the "price of thought" is a fuzzier idea than the price of electricity.

For a business, one caveat outweighs the rest: "essentially no user stays permanently on the cost frontier." Epoch's curve assumes you always pick the cheapest model that can do the job. Hardly anyone works that way. Most people use whatever model their software ships with, at whatever setting it defaults to.

Noam Brown wants the price next to the score

On Wednesday, OpenAI researcher @polynoamial replied to CAIS and Scale AI about HLE-Diamond, their new 1,000-question slice of the Humanity's Last Exam benchmark. Every model on it was tested at "reasoning high," and he pointed out that "high" means very different amounts of thinking from one model to the next. His suggestion: "Why not just report the $ cost of the evaluation?" Even better, he wrote, plot accuracy against dollars.

That's Epoch's point coming from the other side. A benchmark score with no price attached tells you very little about whether a model is worth buying.

Why your bill can still go up

A McKinsey report covered by Fortune on Wednesday explains the catch. McKinsey's Lari Hämäläinen says that as the cost of a unit of intelligence collapses, the amount businesses consume is exploding. Agents make it worse because their costs swing: the same task can cost up to 30 times more from one run to the next, depending on the path the agent takes. We saw a real case of this when Databricks rolled out GPT-6 Astra and its bill went up.

McKinsey's advice is to stop counting tokens and look at the task. What does one run cost, how often does the agent get it right, and how long does a person spend checking the result? Hämäläinen gives an example. If a job takes a person an hour and checking the agent's work takes six minutes, an agent that succeeds just over 10% of the time can already pay for itself.

My read

I believe Epoch's number, and it's good news for small businesses. Work that cost too much to automate a year ago, like reading every invoice or sorting every inbound email, may cost very little now. You only get that benefit if you go back and check your model choices, because the default in your tool usually isn't the cheap option.

I'd watch the total as closely as the rate. A cheaper model that runs ten times as often, or an agent that keeps retrying until it gets lucky, can still hand you a bigger invoice.

What this means for a business using AI

• Price each automation by the finished job, and count the time someone spends checking it. Per-message and per-token numbers hide that.
• Every few months, test whether a smaller, cheaper model can now handle the task. With prices falling this fast, it often can.
• Cap retries and set a monthly budget alert on every agent so one bad loop can't run up the bill.

If you want to know which of your own processes are now cheap enough to automate, that's what our free process audit is for.