Maybe you don’t need the brand name AI

TL;DR: Two cheap open-weight AI models dropped in the last 48 hours: Thinking Machines’ Inkling and Moonshot AI’s Kimi K3 (which is blowing people away). It’s a sign that AI is becoming more of a buyer’s market where consumers can shop around for the best price—and it’s putting the pressure on legacy labs.

What happened: Inkling is the first model from an ex-OpenAI crew led by former CTO Mira Murati. The selling point: While it’s not as capable as the top dogs, it uses just a fraction of its total parameters on a single task to keep it faster and cheaper to run.

Kimi K3, a Chinese model, is more powerful and has been topping some AI leaderboards in testing. It’s the largest open-weight model ever released, with some analysts likening its performance to top US models such as Opus 4.8 (while being about 40% cheaper).

Openly cheap: This may come as a surprise, but the average price of an AI token has actually fallen by about 90% since 2023, as hardware and software improvements made tasks cheaper to run (though high-end models are generally a lot more token-hungry).

It’s easier to take advantage of that price drop with open-weight models, because they can run locally on your own machine or through a variety of third-party cloud providers competing to offer the lowest price. As a bonus, you can fine-tune them more than the closed-weight options from incumbent AI labs, which can also charge a markup because you can only get the models through them.

US labs have a spicier explanation for why some of the most popular Chinese models—many of them open weight—cost so little: They accuse them of distilling proprietary American models to skip the expensive research and training phase.

Uh-oh for US frontier labs?: Companies like OpenAI and Anthropic offer more budget-friendly models, too—but the ones they spend most of their energy promoting are their pricey, cutting-edge flagships (especially ahead of their IPOs). But OpenAI is reportedly weighing big price cuts to draw in more customers, per the Wall Street Journal, and Anthropic has repeatedly kicked the can on switching Fable 5 to costlier usage-based pricing—a sign, perhaps, that it’s worried about attracting enough paying users to its top-tier model.

Bottom line: We’re entering an era where the most attractive AI feature might not be raw capability, but bang for buck. For those who don’t need a top-of-the-line Lambo, there are plenty more good ole reliable Camrys on the lot now. —WK