If the winter of 2026 saw companies gorge themselves on all-you-can-eat AI, summertime is for counting calories.
Prices for some of the most popular tools have risen, and companies that previously asked employees to go all out on AI code are suddenly facing hefty bills. Executives and workers told Business Insider how this new era in AI’s evolution has thrown their workflows and budgets for a loop.
Coinbase executive Rob Witoff is sitting front row on the roller coaster. After Anthropic’s Claude launched its much-improved coding model Opus 4.6 in February, Witoff, who oversees the crypto exchange’s infrastructure, said, “Our internal usage started to go parabolic across the company.”
Now, with prices rising, Coinbase has instituted a sophisticated system of weekly price caps ranging from $500 to $5,000 based on each employee’s job level and role.
“Once people understand what’s possible, usage takes off on its own. Then the focus shifts from ‘Are people using AI?’ to ‘Are they using it well?’ That’s where we are today,” he said.
It’s the latest whiplash on the AI frontier. Executives and developers are rethinking their rapacious demand of recent months by switching models, imposing limits, and prioritizing projects. As a result, AI juggernauts, including OpenAI and Anthropic, risk losing market share to cheaper models — at a time when they’re racing toward stock-market-altering IPOs.
The high stakes are prompting scrutiny. Salesforce CTO Parker Harris told Business Insider that the company has fully opened the floodgates for spending on Anthropic tools, but that likely won’t last forever. They’ll have to find a balance that doesn’t divert too much money to the rising startup.
“We gotta run a business, we’re a public company,” Harris said. “We can’t tell our investors like, ‘Yeah, sorry, we gave half of our upside this year to Anthropic so they can go public’.”
Tokenmaxxing is out, restraint is in
Between February and June, OpenAI, Anthropic, and GitHub each shifted their pricing models. One by one, the companies began charging more customers based on their token usage — the units that measure AI’s input and output — rather than with flat-rate billing.
The corporate world quickly changed course. Walmart placed usage limits on its internal programming tool. Amazon shut down an internal “tokenmaxxing” leaderboard. Accenture, IBM, Oracle, and JPMorgan Chase have backed a new “Tokenomics Foundation,” meant to standardize AI budgeting metrics across companies. By April, Uber had already blown through its AI budget for the year, and executives said its glut of token spending hadn’t translated immediately into useful releases.
Leaders have begun to view token waste as fiscally irresponsible, Niranjan Krishnan, the head of AI solutions at the IT consultancy FPT Americas, told Business Insider.
“The novelty has worn off, and hard-nosed utility has stepped in,” Krishnan said. “That’s 2026 for you. The magical thinking era is gone.”
A March and April survey of 200 executives by Wakefield Research, on behalf of the AI accountability startup Lanai, found that 79% of respondents were slightly or very concerned that their AI budgets would be cut because the spending wasn’t tied to new revenue or profits.
A senior software engineer at Deloitte said that the changes to GitHub’s pricing model are “already wreaking havoc” on expectations for work, with developers quickly burning through their new monthly quotas, which took effect in June. The software engineer estimates that a single, highly detailed prompt, which could have a model working for hours, would now cost more than $100 under GitHub’s new usage-based billing.
“The cheap ‘AI buffet’ days are over,” the software engineer said. Developers will now need to use AI tools more deliberately, narrowing their prompts and breaking large jobs into smaller tasks instead of handing autonomous agents sprawling specifications to complete on their own, he said.
Coinbase’s Witoff pointed to an extreme hypothetical example of AI use that may have been fair game in the olden days of winter, but would now require much more scrutiny: analyzing all of the company’s code for bugs using one of the state-of-the-art AI models.
“At our scale, that might cost $50,000 to $100,000 a run,” he said. “So if you’ve got a hundred people doing that independently, you’re going to spend $10 million.”
Some projects are worth it. Coinbase’s system alerts users when they’re approaching their limits, but they can apply for an exception, which Witoff said are often granted. The system is meant to make workers more mindful about how they’re using tokens.
“We think constraints breed creativity,” Witoff said. “We don’t want people burning money just because they can, or creating the wrong incentives.
LogicMonitor is also imposing internal limits on token use and including them in some products it sells to customers. The IT company’s head of AI, Karthik Sj, told Business Insider he is now less likely to leap to the priciest, most powerful AI tool to do a task, unless it’s a marked improvement above the one they were using prior.
“We are in uncharted territory, and I think this is going to give a reckoning moment for many companies, many CIOs, many CFOs,” Sj said. “How do we not dread this tokenmaxxing situation, and really focus on value? It’s a good problem for the industry, I think.”
Why AI subsidies had to go
OpenAI’s CEO, Sam Altman, has heard the groundswell of concern. He said at a recent event that AI budgeting went from “at the beginning of this year, an issue that never came up — people were totally happy with the amount they were spending — to all of a sudden, a huge issue.”
So why did prices rise? For OpenAI, Anthropic, and GitHub, the economic proposition has changed. Tokens have become cheaper thanks to Nvidia’s chip innovations, but AI tools’ popularity and new agent-based setups meant providers could no longer afford to subsidize heavy users.
Mario Rodriguez, GitHub’s chief product officer, wrote in a post announcing the new pricing that under the company’s old billing setup, a “quick chat question and a multi-hour autonomous coding session can cost the user the same amount.” He said GitHub had been eating the costs, but it was “no longer sustainable.” GitHub’s move was met with a wave of complaints from developers, as some learned the shift to usage-based pricing would multiply their AI code expenses to thousands of dollars a month.
OpenAI and Anthropic have been taking steps to alleviate costs, promoting their newest model releases as more “token-efficient” than prior models. Both labs offer an option to complete non-urgent tasks slowly but at a lower price. They also offer prompt caching, which saves previous queries to use as reference points, cutting down on required computing power.
An Anthropic spokesperson told Business Insider that the company sees its new usage-based model as benefiting customers because it’s more customizable — heavy users aren’t cut off, and light users don’t pay for capacity they don’t use. OpenAI and GitHub did not respond to Business Insider’s requests for comment.
Companies are making their own cost-cutting fixes. At the software startup Harness, use of Claude Code drove AI costs to “grow exponentially” from October through March, senior vice president Trevor Stuart told Business Insider. By training engineers and building internal tools, the company has managed to reel in costs over the last few months. Because of the pricing changes, Stuart said, “I think many teams are now having that same conversation.”
All-you-can-eat out, regular ol’ menu in
The AI coding craze, for all the twists and turns of 2026 so far, is still in its infancy. The “tokenmaxxing” uproar, the sticker-shock pricing, the see-sawing behavior by companies — these are all symptoms of a sudden paradigm shift, and companies are proceeding with caution.
Salesforce’s CTO Harris said the company is spending “far more” than they’d planned on tokens for its 2026 fiscal year. For now, he said, they’re seeing where the change takes them, trying to avoid “holding back on the gas.”
In the meantime, Salesforce is rolling out measurements for both customers and internal engineers to better understand the tangible impacts of AI. Harris noted that the company could have been measuring software engineers’ output more closely before, and said its new code-focused metric, called an Effective Output score, will help avoid further surprises.
“We have enough data now that we can forecast,” Harris said. “And then it’s just a question for the company: ‘What’s the return? What’s the right spend for the return?'”
That calculus has some executives angling for cheaper AI. Some companies Business Insider spoke to, including Coinbase, have begun offloading basic work to less advanced AI models, either from American companies or from Chinese firms like Deepseek and MiniMax.
Ahmad Awais, the founder of the coding agent startup Command Code, told Business Insider his company gained 10,000 customers in a recent 30-day stretch, largely driven by demand for cheaper models.
For some of those businesses, using the most advanced — and expensive — models isn’t worth it for everything. Stuart, the Harness executive, likened using a cutting-edge AI model for basic text-summary work to “taking the Ferrari to the grocery store.”
Have a tip? Contact AI reporter Stephen Council via email at [email protected], or over text, Signal, Telegram, or WhatsApp at 415-757-8198. Use a personal email address, a nonwork WiFi network, and a nonwork device; here’s our guide to sharing information securely.
