Most software program comes with a reasonably predictable price ticket.
If your organization buys 1,000 Microsoft Workplace licenses or rents computing energy from Amazon, you possibly can usually estimate the invoice.
However AI brokers are totally different.
An agent may spend hours researching a subject, writing software program or finishing sophisticated assignments. It’d want to look information, name exterior instruments and make dozens and even lots of of requests to an AI mannequin.
And each a type of steps consumes computing energy.
In different phrases, the extra work we hand over to AI, the extra intelligence we’re going to devour.
And as we talked about yesterday, the value of that intelligence can differ enormously relying on which AI mannequin you utilize.
However a brand new examine uncovered one other wrinkle that might make managing these prices much more tough.
As a result of even when you understand the price of the mannequin you’re utilizing, you continue to may not know what the precise job will value.
And as companies start deploying AI brokers at scale, that uncertainty can have huge implications for the economics of AI.
The Agent Tax
The examine I discussed comes from researchers at Stanford and a number of other different universities who needed to reply a easy query:
How a lot does it really value to place an AI agent to work?
They examined eight main AI fashions on real-world coding assignments, monitoring the tokens every one consumed. Tokens are the small items of knowledge an AI reads and produces.
And since AI corporations typically cost based mostly on token use, extra tokens imply an even bigger invoice.
These weren’t easy questions with one-shot solutions. The brokers needed to work by way of software program issues very like human programmers, studying information, deciding what to alter, testing their work and attempting once more when one thing went fallacious.
That’s the place the primary shock appeared.
On common, an AI agent engaged on a coding task used roughly 1,200X extra tokens than a coding chat. And it used 3,500X extra tokens than an AI answering a coding query in a single response.
Most of these additional tokens didn’t come from the solutions the AI produced. They got here from the knowledge the agent repeatedly fed again into the mannequin because it labored.
That’s as a result of every time an AI agent takes an motion, it’d must reread a lot of what has already occurred so it is aware of the place it’s within the activity.
That’s how a comparatively easy job can flip into an unlimited quantity of AI utilization.
However the researchers discovered one thing much more shocking.
After they ran the identical activity utilizing the identical AI mannequin a number of occasions, token consumption may differ by as a lot as 30X.
That’s as a result of AI brokers don’t all the time take the identical path to a solution. One run may discover the best answer shortly. One other may spend for much longer exploring useless ends earlier than ultimately reaching the identical place.
However what I discover much more fascinating is that spending extra tokens didn’t essentially produce higher outcomes.
As these charts from the examine present, accuracy typically peaked at a reasonable degree of token use after which stopped enhancing. In some instances, it even declined.
So the additional computing energy was basically being spent on extra looking out and reasoning that didn’t make the ultimate reply any higher.
The variations between fashions have been additionally substantial.
On the identical assignments, Kimi-K2 and Claude Sonnet 4.5 consumed, on common, greater than 1.5 million extra tokens in contrast with GPT-5.
What’s extra, when researchers requested the fashions to foretell what number of tokens they would wish earlier than beginning a activity, they couldn’t do it very nicely.
Even the most effective fashions confirmed solely a weak relationship between their estimates and what they ultimately consumed. Total, they tended to underestimate their very own prices.
This examine centered particularly on coding brokers. However the identical fundamental downside may emerge each time an agent works by way of an extended, sophisticated task.
And this implies companies may not know what an agent will value till after the work is completed.
That’s not how companies favor to function. As I wrote about earlier this month, Uber just lately discovered this out the laborious manner.
However Uber isn’t alone.
Microsoft just lately started telling its personal engineers to pay nearer consideration to what number of AI tokens they devour. The corporate is setting token budgets for groups and steering staff towards cheaper fashions when probably the most highly effective ones aren’t essential.
And Atlassian has launched month-to-month AI “wallets” for its builders, starting from $500 to $2,000 per thirty days.
Its CEO warned that autonomous brokers operating within the background can devour huge numbers of tokens with out workers essentially realizing how shortly the invoice is rising.
Right here’s My Take
AI intelligence retains getting cheaper.
However worth is just half the equation.
An inexpensive mannequin that takes 20 steps may finally value greater than an costly mannequin that solves the identical downside in two. But an agent that saves an worker an hour of labor may nonetheless be a discount, even when it consumes hundreds of thousands of tokens alongside the way in which.
So companies must begin interested by AI when it comes to each value and effectivity.
And that’s the place the economics of AI will get fascinating.
As intelligence will get cheaper, we’re going to search out extra issues for AI brokers to do. And the extra work we hand over to them, the extra computing energy they’ll devour.
Which suggests the falling value of AI may finally lead us to spend extra on it, not much less.
And that might maintain the AI infrastructure growth operating for lots longer than Wall Road expects.
Regards,
Ian King
Chief Strategist, Banyan Hill Publishing
Editor’s Observe: We’d love to listen to from you!
If you wish to share your ideas or options in regards to the Every day Disruptor, or if there are any particular matters you’d like us to cowl, simply ship an electronic mail to dailydisruptor@banyanhill.com.
Don’t fear, we received’t reveal your full title within the occasion we publish a response. So be happy to remark away!
