Can tech companies learn to love cheaper AI models? 

The AI growth has been constructed on a fundamental assumption: Greater fashions are extra highly effective, and probably the most highly effective fashions win. Now, the trade is about to be taught what occurs if that assumption begins to interrupt.  

Mounting prices have already pressured customers to present smaller and cheaper fashions a re-evaluation. This cost-conscious model-shopping is new and it’s unclear the way it will have an effect on the trade, however the influence is more likely to be vital. 

One prediction, laid out greatest by Coinbase co-founder Brian Armstrong, is that it’ll consequence within the overwhelming majority of duties shifting to cheaper fashions. 

“[D]emand for intelligence is close to infinite, however 80% of workloads shall be operating on 99% cheaper fashions inside 12-18 months,” Armstrong wrote on X. “20% of workloads will nonetheless run on newest gen fashions the place IQ maxing is necessary.” 

It’s laborious to overstate what a big shift it will likely be for the AI trade if Armstrong’s prediction comes true.  

Prior to now, most AI firms have competed on high quality, which has meant defaulting to probably the most superior accessible mannequin. If those self same jobs might be dealt with by cheaper fashions with out affecting high quality, it could imply an enormous shift within the economics of AI. And critically, a lot of the financial savings could be popping out of the pockets of the large labs, dealing a monetary blow to OpenAI and Anthropic simply as they’re heading for their IPOs. 

It’s a probably seismic change within the trade, resting on one fundamental query: Are firms prepared to change to smaller fashions? 

Preliminary assessments recommend that, when the system is organized proper, cheaper fashions may sub in with none sacrifice in high quality. In a latest check by the authorized AI software Harvey, the firm was in a position to scale back inference prices by 3x with out decreasing high quality. The check, carried out in partnership with the inference platform Fireworks AI, mixed Claude Opus and Fireworks’ GLM 5.1, and shifted to Opus for probably the most intensive duties. The consequence was a considerably decrease load when it comes to server time and general price. 

“High quality comes first, and in authorized it at all times will,” Harvey co-founder Gabe Pereyra instructed Trendster, referring to the AI authorized companies his startup supplies. “Nevertheless, the definition of high quality is evolving from merely utilizing probably the most highly effective mannequin for every thing, to utilizing the perfect mannequin that will get the correct reply most effectively.”

This development is commonly framed when it comes to main labs versus Chinese language fashions or open-weight ones, however that misses the larger level. The true divide isn’t between proprietary and open fashions; it’s between giant fashions and small ones. You can get monetary savings by switching from GPT-5.5 to DeepSeek’s V4 Flash, however switching to GPT-5.4-mini works simply as effectively.  

There’s an lively value warfare occurring between in-house inference from the large labs and independently served open-weight fashions. For the larger query of small versus giant, it doesn’t actually matter which sort of small mannequin wins out.  

All of this may appear apparent — after all you shouldn’t use extra compute than obligatory — but it surely runs counter to the scaling-first strategy that has dominated the trade till now. Impressed by the bitter lesson, labs have leaned laborious into coaching probably the most compute-intensive fashions doable, pushing the frontier of what AI fashions can do. With costs closely sponsored by traders, purchasers had no motive to decide on something however probably the most superior choice.

With token costs rising and subsidies slowing down, customers are dealing with price strain for the primary time. We don’t know whether or not the brand new price strain will really drive enterprise customers to smaller fashions. They may simply as simply economize by making fewer calls, utilizing much less context, or just giving up on the least promising deployments. 

But when it seems that almost all deployments might be run simply as effectively on a smaller mannequin, it may put a critical damper on the rising demand for inference — and lift new questions on how you can justify the price of coaching a frontier mannequin.