• Anthropic's Boris Cherny says companies are right to focus on their ROI for AI.
• At the same time, Cherny said employees at all levels and roles still need tokens to be able to experiment with AI.
• Then, the Claude Code creator said, companies can start to control costs.

Claude Code creator Boris Cherny has a message for companies that are nervous about their AI token budgets.

"ROI is absolutely the right framing because you don't want to just think about cost because you kind of spend something on it and you get something back," Cherny said during a recent fireside chat at Scale AI.

Jesse Chen, Meta's director of product management who moderated the chat, asked the Anthropic employee directly about the recent concerns raised by Uber COO Andrew Macdonald about whether the rideshare giant's AI spending was leading to enough of a return to justify the rising cost of AI tokens.

Tokens are units of text that serve as a measurement for AI usage, such as the prompts processed by large language models, including those that power chatbots like Anthropic's Claude or its generative AI coding tool, Claude Code.

Cherny said it's right to be focused on ROI. It's also important, he said, not to overdo it in response to cost concerns.

"The way to do this is give people tokens and give them safety to experiment so they feel like they can try stuff and they're not going to get penalized for it," he said. "Once you find these internal use cases that kind of work, then you want to control the costs and you want to do that on the backend, not on the front end."

Otherwise, companies might miss out on the best ideas for deploying AI.

"Often, some of the most interesting ideas and the most innovative ways to improve processes and new product ideas are going to come from an accountant somewhere in the corner of the org or a marketing person that the CEO has never heard of," Cherny said.

Cherny emphasized that Anthropic offers several ways for its enterprise customers to control costs and set budgets, including per-seat cost controls.

Others in the AI space, including OpenAI CEO Sam Altman, are also increasingly discussing companies' concerns about the ROI of their AI investments.

As Cherny mentioned, AI firms like Anthropic are essentially token generators. That also means that they have an incentive to keep selling their models and generative AI tools, especially as they approach highly anticipated IPOs. The creator of Claude Code said that Anthropic is also paying attention to how its tokens are used.

"They're not free for us because every token we use is a token we do not give to a customer, so there's an opportunity cost," he said. "When I think about it, it actually maybe comes back to ROI."

Measuring that ROI is also changing, Cherny said, as the pace of AI model advancements continues to accelerate. He previously said that companies may have looked at the percentage of code written by AI. Cherny said that measurement is no longer as useful once more people let AI write 100% of their code, as he does.

"Then think about, how much is the code per engineer accelerating? And then the third thing to think about is like, what are the other bottlenecks that are getting in the way?" he said. "Because once you get it to this point where engineers are just writing a lot of code, the bottleneck is going to be like good ideas. So, how do you un-hobble that so that your company can generate ideas faster?

• CFOs are taking charge of AI spending as companies pour billions into the technology.
• Some are introducing AI budgets and new controls to keep costs from spiraling.
• "The CFO is really becoming the face of the AI story," said a PwC advisor to finance chiefs.

At Match Group, every employee now has an AI budget.

The parent company of Tinder, Hinge, and other dating apps recently began giving department heads a set amount to spend on AI, which is then distributed across their teams. Employees can track their usage on a dashboard, and if they want to exceed their budget, they have to explain why. The company's most expensive AI models also aren't available by default and require a specific use case.

"If you don't set guardrails, there's no reason for an engineer to not go use the most expensive model," said Match Group CFO Steve Bailey. The average software engineer at the company spends roughly $600 a month on AI tokens, he said.

Match Group's system reflects a growing reality across corporate America: As companies spend billions on AI, CFOs are emerging as some of the most powerful executives in the AI era.

Finance chiefs are doing more than signing off on AI budgets. In many cases, they're the ones deciding who gets access to AI tools, how much employees can spend, which vendors make the cut, and whether AI investments are generating enough value to justify costs.

"The CFO is really becoming the face of the AI story," said Peter Pollini, a PwC advisor to finance chiefs in the financial-services sector.

A spending boom

The stakes are enormous. Match Group initially allocated $5 million for AI this year, but it's now on track to spend double that amount, Bailey said. The increase followed CEO Spencer Rascoff's May push to make the company more AI-native by expanding access to AI tools across the workforce. Initially, they were available mainly to engineers.

"Aside from maybe travel and entertainment, we've never had to budget for a cost that's this big at the employee level," Bailey said.

To help fund those investments, Bailey said Match Group plans to dramatically slow hiring while it assesses how AI could reshape its workforce.

Across corporate America, similar calculations are turning CFOs into the gatekeepers of one of the biggest spending booms in decades.

At Elevance Health, CFO Mark Kaye oversees a hidden way of keeping AI costs from spiraling. The insurance giant quietly routes employees' queries to different AI models based on the complexity of the request. That's because a single prompt can cost anywhere from a few pennies to more than a dollar, depending on how many tokens, or units of data, employees gobble up.

"We manage it on the back end," said Kaye, adding that he expects Elevance Health, the parent of Anthem Blue Cross Blue Shield, to invest $1 billion or more on AI this year.

Making AI pay off

Some CFOs are making tough decisions about how to fund AI expenses at large, such as by freezing annual salary raises and laying off workers. Others say AI is helping to pay for itself.

Kaye said AI automation at Elevance has reduced administrative work tied to medical-chart reviews by roughly 40%, giving staff more time to support customers.

"There are significant inefficiencies in the system that AI is allowing us to take out," he said.

Keeping a tight leash on AI spending isn't the only new hurdle for CFOs. They're also responsible for managing spending on a category that is evolving more rapidly than previous generations of enterprise software.

For the first time this year, Xero, a global small-business platform that offers accounting, payroll, and payments, added a line item to its budget for AI token spending per employee, said Claire Bramley, the CFO. The company also created a task force to review software purchases and identify AI products it can do without.

"Do we have more than one tool that serves the same purpose?" Bramley said. "As a CFO, you want to make sure that everybody's not going off and doing their own thing."

AI is also changing who CFOs spend time with. Bramley said finance, technology, and HR leaders at Xero now work together more frequently to evaluate software purchases, hiring plans, and how AI could affect future staffing needs.

"You could probably do it once a month before, and I think you have to do it weekly today," she said.

Additional headaches

CFOs are also facing new business problems arising from AI.

Netta Samroengraja, finance chief at healthcare platform Zocdoc, said her team has had to hustle to evaluate AI tool providers to solve problems that, ironically, were created by the technology. In recruiting, for instance, the technology suddenly enabled job seekers to flood the company with applications and create phony personas.

"It was pretty prevalent very quickly, and so we had to react quickly," Samroengraja said.

That wasn't the only surprise, as the economics of AI were shifting, too. Early on, Zocdoc raced to vet vendors, anticipating that prices designed to attract customers at the start of the AI boom would increase over time.

The company used that window to test multiple providers and compare their cost and effectiveness before settling on the tools that delivered the strongest business results, Samroengraja said, adding that Zocdoc has been willing to spend more on tools that produce measurable business outcomes rather than optimize for the lowest possible AI spend.

"If you see the ROI in it, you should keep investing in this," she said.

A crowded AI market is making those decisions even harder. New providers are constantly pitching tools that promise to boost productivity, cut costs, or replace existing software, forcing many CFOs to take a more active role in evaluating vendors, said Alex Sobol, cofounder of the Millennium Alliance, an invite-only community for C-suite executives in North America and Europe.

"It seems like every hour there's a new AI vendor," he said. "It's hard to know what's real and what's fake, and what's good and what's bad."

• Tokenmaxxing is all the rage in Big Tech. For startups, the trend is opening up debate.
• Some founders told Business Insider that they spent big on tokens; others used capped subscription plans.
• One founder called tokenmaxxing "extremely stupid." Another said: "You've got to spend money to make money."

Kavitta Ghai wants her startup's engineers to spend more tokens.

The 29-year-old cofounder of Nectir started setting minimum quotas for Claude Code use. First it was at least $100 in tokens a week, then $200. Now, the expectation is that her engineers each spend a couple thousand in AI tokens a month.

The strategy has been successful, Ghai said. Some of Nectir's senior engineers were previously skeptical of AI coding tools; now, they call it their "army of coders," she said.

But she doesn't think Nectir is "tokenmaxxing," the buzzword du jour for techies racing to spend as much as they can. "We don't really play into the Silicon Valley trends," Ghai said. "We live in our own world, and we're competing against ourselves."

Across Big Tech, engineers are racing to spend as many tokens as possible. A token is a measure of AI compute. The more tokens burned, the more the engineer employs AI tools. Employees at Meta reportedly competed on a token leaderboard before it was taken down.

What of the little guy? Startups are an edge case: relatively tiny teams that want to be on the cutting edge of tech but might not have the same money to spend. Some startup leaders told Business Insider that big token bills helped them succeed. Others scoffed at the idea, preferring to stick to the lower-cost subscriptions.

The startups spending big on tokens

Aron Solberg doesn't want the competition of a token leaderboard — but he does want the mindset behind it.

The 44-year-old cofounder of Risotto sees token spending as a "force multiplier" for a small team. The company uses OpenAI and Anthropic's models, and said it spends $4,000-5,000 per month on tokens. Six months ago, Risotto says he spent one-tenth of that sum.

"It's trending up a lot," Solberg said.

"There's an old adage that rings true," he said, whether it was for hiring new employees or spending liberally on tokens: "You've got to spend money to make money."

Quang Hoang is similarly spending big. He wrote in an email that his startup, Vybe, has an "unlimited credit policy" and was thinking about minimum quotas.

Investors are also incentivizing spending — and might foot the bill.

Hoang tells founders he invests in to allocate "at least their salary amount to tokens." (Nvidia CEO Jensen Huang made headlines last month for saying he would be "deeply alarmed" if one of his $500,000 engineers did not consume at least $250,000 of tokens.)

Accelerators like Y Combinator offer free token credits to their participants. "At YC, we let our engineers let it rip," CEO Garry Tan wrote on X. Those credits help some founders to spend big. These founders aren't tokenmaxxers, but do believe that there are productivity benefits.

Traverse cofounder Lance Yan believed in Tan's message: "We usually just let it rip." The 19-year-old said he uses the best models with the maximum effort, not worrying about the costs. Between his Claude Max subscription and the credits that offered by YC, he can spend big without hitting a limit.

He's not a fan of rationing tokens. "That's stupid," he said. "You're just harming your own startup."

26-year-old Boris Skurikhin said that the YC credits helped his startup Docket get off the ground. He's mostly run through them now, except for the models he uses less frequently.

Skurikhin said he noticed a 10x increase in productivity in his own work when he used the tools. "It is expensive to build with tokens," he said, but "not as expensive as having another engineer."

Many of these startups are in the AI game, after all. Nectir's Ghai said that token spending instilled "AI literacy" — something that's especially important, given their product.

"The team itself needs to be the best versed at it first, before we try to go sell it to anyone else," she said.

The startups saying no to tokenmaxxing

Rishabh Sambare wishes he could spend more on tokens.

The 23-year-old cofounder of Gale prefers to build with Zed, an AI IDE similar to Cursor, but can't stomach the company's usage-based pricing. The subscription deals from OpenAI and Anthropic are so deeply subsidized that he uses them instead.

"It sucks, because I hate their products," he said, calling Zed "more polished and less buggy between releases."

Sambare is Gale's only engineer, though the company often has 2-3 interns. He hasn't hit a rate limit, but one of his interns has. They got him a second subscription, he said; it was still far cheaper.

These subscriptions — sending $100 to $200 to Anthropic for its "Max" tiers or $100 to OpenAI for its "Pro" plan in exchange for a stable of discounted tokens — were popular among the founders I spoke to. Hassan Ismail, the 24-year-old founder of Argos Research, said the Claude Max subscription was a "no-brainer," and that all five team members have a $200 a month subscription.

Others were more philosophically opposed to the trend. Weave's Brennan Lupyrypa didn't mince his words: "It's extremely stupid for any company to be tokenmaxxing."

Weave is still spending big on tokens because it doesn't want to "kneecap" its engineers, its 25-year-old founding engineer said. The company set up a notification for when an engineer hit $500 in token spending a month; Lupyrypa said most hit it within two weeks.

But Weave doesn't incentivize the spending itself, which Lupyrypa said was the wrong proxy. He predicted the downfall of tokenmaxxing within the next three months. "CFOs won't be happy," he said.

Still, some tokenmaxxers hold strong. I asked Risotto's Solberg about these token-hesitant founders. He said that they likely hadn't found their product-market fit yet.

"It makes complete sense to spend a lot of money on tokens, because you know that the growth is coming soon after," Solberg said. "If you're a venture-backed business, that's what you signed up for."