As some companies pump the brakes on AI spending by resorting to cheaper models, others are going all-in on the most advanced AI systems -- even with their hefty price tags.
So-called frontier AI models, or the most capable systems made by companies like OpenAI and Anthropic, can be expensive to use partly because they require a lot of compute and process large numbers of tokens, AI's basic unit of measurement. But these state-of-the-art models are considered the best because they can " reason" through complex, multistep problems and are capable of supporting a variety of tasks, including powering autonomous AI agents.
The calculus often is as much a business decision as an engineering decision. If paying a premium for a frontier model means a better product or an upper hand over rivals, many companies say it's worth it.
At Shopify, for instance, engineers aren't allowed to use less powerful "non-frontier" AI models, says Farhan Thawar, vice president and head of engineering at the e-commerce company. Shopify uses Anthropic's Claude Code, as well as coding agents and models from other AI providers, he said .
"I don't really want them to use something smaller, typically because you end up losing human time versus computer time," Thawar said. If an engineer opts for a smaller model, and that model misses something a frontier model wouldn't have, then "the human time to figure that out is not worth the toil," he said.
"I typically am not as worried about token cost because I'm learning faster than I would have without the tokens," he added.
Bill Nguyen, a Silicon Valley veteran who recently founded an AI voice startup called Olive, used a startling 774 billion tokens in the last month and a half, mostly for his personal use, he said. That's equivalent to roughly $4.5 million in computing, he added.
Nguyen says he's mostly using frontier models because the smaller and less advanced models simply don't work as well. "You'll pay the money for the difference because you get way better outcomes," he said.
In his view, startups and companies that are challenging established businesses need the frontier models to gain an edge. In other words, in the race to build the next, better product, you'll get there faster with frontier models.
"If time to market and competitive risk is more important, and you're willing to spend for it, there's no way you'll choose anything other than a frontier model," Nguyen said.
The cost is probably not worth it for simple queries and tasks like summarization and editing, where the difference between frontier and cheaper models is negligible, tech leaders and analysts say. Indeed, there is an ongoing debate over whether AI models are becoming a widely-available commodity.
But for complex reasoning tasks like managing AI agents, advanced coding problems and multistep research, frontier models perform better -- even if by a small percentage -- and that can make all the difference in outpacing the competition.
In accuracy, frontier models gained nearly 30 percentage points -- going from under 10% to roughly 38.3% -- in a single year on Humanity's Last Exam, according to Stanford University's Institute for Human-Centered Artificial Intelligence. The benchmark exam is designed to evaluate whether AI can solve problems that would be difficult for highly-educated humans.
Other companies say they're choosing frontier models because they need top-of-the-line capabilities.
Tyson Chen, co-founder and co-CEO of Avoca AI, said the startup generally sticks with frontier models. "The workflows we're delivering are quite revenue-sensitive, so every percent of accuracy is important," he said.
Still, the startup does use "cheaper alternatives" in some of its products where latency -- the time a system takes to respond -- is less critical, he said.
Avoca isn't alone. With the cost of AI rising, more companies are using smaller, cheaper models, and open-source or open-weight models. It has also become more popular to use cheaper models for less critical tasks -- allowing companies to save on token costs.
At companies like Spotify, it's an ongoing discussion whether frontier models are worth the cost.
Niklas Gustavsson, Spotify's chief architect and vice president of engineering, said the audiostreaming platform decided against giving its engineers the latest two versions of Anthropic's most advanced Opus models.
Gustavsson said those models' performance improvements don't justify "the higher costs that they happen to have. But these are the exact types of discussions we're having at the moment."
The dilemma for companies like Spotify is that every release of a new frontier model sparks yet another evaluation of model capability versus cost. Gustavsson says he's seeking to balance the two, while equipping engineers with the information they need to make decisions on which model to use for certain tasks.
Developers tend to love using frontier models because "they simply work better," said Philip Walsh, an analyst focused on software engineering at market research and IT consulting firm Gartner. But most companies are trying to find ways to make sure workers use more cost-efficient models or are building AI agents that can take advantage of frontier models for planning tasks, while relegating lower-tier tasks to cheaper models, he said.
Khozema Shipchandler, chief executive of Twilio, said the customer-engagement software company is using a variety of AI models, but hasn't deployed any open-source models even though "the economics are compelling."
"The big question every company, including Twilio, has to reckon with is: Are we truly driving ROI [return on investment] with our AI usage? There will come a time when the concept of 'tokenmaxxing' will be remembered as completely reckless," he said.
Part of the way Anthropic has tried to address token costs is by offering customers a variety of tools to control costs.
Boris Cherny, the head of Anthropic's Claude Code, said the AI lab allows customers to put spending limits in place and opt for some of Anthropic's lower-cost models. Customers can also "tune" how much thinking a model does -- essentially asking a model for less intelligence, which uses fewer tokens, he said.
"It's just keeping costs reasonable and predictable," Cherny said. "But I think actually the far bigger opportunity is increasing return, and I think this is what customers are saying, too. The more tokens that they use in a useful way, the more return they get."
Write to Belle Lin at belle.lin@wsj.com
(END) Dow Jones Newswires
July 17, 2026 16:09 ET (20:09 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments