The AI industry has spent the last few years obsessed with one question: who has the smartest model? That still matters. But a new north star is emerging: how much useful intelligence can be delivered for every dollar spent. Many models are now good enough for many commercial tasks. Once that happens, buyers start to
The AI industry has spent the last few years obsessed with one question: who has the smartest model?
That still matters. But a new north star is emerging: how much useful intelligence can be delivered for every dollar spent.
Many models are now good enough for many commercial tasks. Once that happens, buyers start to care less about the best model and more about the cost of doing reliable work with AI.
Eugene Kim’s recent scoop on Amazon’s rebuild of Alexa+ is a good example. Internal documents show that the company directs more requests to its own less powerful AI, while avoiding unnecessary calls to Anthropic’s more expensive and higher-performing models.
The goal was not to make Alexa always use the smartest model. It was about using expensive intelligence only when the job required it.
“This is a very strong and real trend,” said Kylan Gibbs, CEO of Inworld, which develops powerful voice artificial intelligence. “We’re getting to a state where many models are good enough, and in that context it’s all about efficiency.”
His company has created independent research teams focused on making Inworld models cheaper and faster to run, not just smarter.
So what is the best AI model, per dollar of intelligence? This is harder to answer than when the industry focused on pure performance. Still, Peter Gostev, AI capabilities lead at Arena AI, shared four things to consider:
- How good is it? How reliably does the model complete the real work?
- What do you charge? Compare the cost of reading a request and producing a response. Some model providers charge more for especially large jobs.
- How much can be reused? Reusing already processed information can drastically reduce the bill.
- How much work is needed? A model that is cheap per token may still cost more if it requires additional steps or repeated attempts to finish the job.
Peter was reluctant to share a clear classification of AI models based on these criteria, in part because this trend is so new.
However, this is becoming clearer: the smartest model may still win the headlines, but the model that offers the most useful work for the money will win the market.
Subscribe to BI’s Tech Memo newsletter here. Contact me by email at abarr@businessinsider.com.