A looming crisis could limit some of your favorite AI tools

• Some AI companies are showing major signs of stress.
• The rise of agentic AI is stretching already finite compute.
• Anthropic in particular is struggling to keep up.

AI companies are getting a brutal reality check, and so are their customers.

As users devour AI tokens, a few frontier AI model makers and hyperscalers are showing signs that they're the ones being maxxed out.

On Monday, Microsoft's GitHub Copilot said it was pausing new signups for its Student, Pro, and Pro+ plans. It is also tightening its usage limits.

On Tuesday, Anthropic said it was experimenting with whether to stop offering Claude Code, its most popular tool, to its lowest-tier paid subscribers. Anthropic later said it was a "test" and that it would inform users before making such major changes.

These aren't isolated incidents; they're warning signs that getting the AI fix users crave will become harder to come by.

Customers are feeling the pain

AI companies hooked users with fast, useful tools that were readily available. In the last few months, there's been more moaning about throttling and soaring costs, and worries about what's to come.

"We get used to these incredible tools only to have them taken away, replaced by a worse or stranger model, and then eventually become outpriced," a user wrote on the Anthropic subreddit.

Increasingly, company representatives are stating that agentic AI has outstripped their initial plans. Thanks to OpenClaw and similar tools, power AI users have found ways to run AI models around the clock. As a result, companies' already limited amount of available compute is being taxed in ways it wasn't six months ago.

"Long-running, parallelized sessions now regularly consume far more resources than the original plan structure was built to support," Joe Binder, VP of Product, wrote in a blog post for GitHub announcing its new limitations.

Binder said that "it is now common" for some AI requests "to incur costs that exceed the plan price."

These aren't the economics that AI companies signed up for.

"It's almost impossible for these companies to build a successful business model with the way they started in 2022," Arun Chandrasekaran, a distinguished vice president analyst at Gartner, told Business Insider during a recent interview.

Chandrasekaran said companies "are trying to wean users" away from free subscriptions but showing the tangible value that the latest AI models can create.

Anthropic is facing added pressure to meet agentic demand after its products experienced an unexpected surge when users flocked to the company following its standoff with the Trump administration.

"Claude literally went from nowhere to being No. 1 in the Apple App Store a month ago," Chandrasekaran said. "The point is that level of sheer demand is not something that any company will be prepared for."

Since the Trump administration effectively blacklisted Anthropic in late February, Claude has experienced multiple outages, and the company has tweaked its usage limits during peak hours.

On Tuesday, Amol Avasare, Anthropic's head of growth, wrote on X that the AI startup's subscription plan wasn't simply built with agents in mind.

"When we launched Max a year ago, it didn't include Claude Code, Cowork didn't exist, and agents that run for hours weren't a thing," Avasare wrote. "Max was designed for heavy chat usage, that's it."

This situation is unlikely to change soon.

Chandrasekaran said compute has regional hurdles that data centers of old didn't have to contend with. US users may not face that crunch as much, but users in other countries might.

"If I'm a user sitting in Belgium, for example, I'm likely to hit the Amsterdam data center, which means that the providers have to deliver that compute capacity within that specific cloud region, within that specific country," he said. "So, it's not like a global pool as well, and that often compounds the problem."

Faced with finite compute, Chandrasekaran said companies will try one of three strategies: make their models more efficient, become better at routing requests, or decide which users to prioritize.

Consumers are unlikely to love any of those options.

Making models more efficient and improving routing means that Big Tech and AI companies will eventually try to deemphasize which particular model is in use, Chandrasekaran said. It also means charging more to access older models, which by their nature are less efficient and thus consume more compute to run.

In the past, companies like OpenAI have sunset older models, sparking outrage among some users. In August 2025, OpenAI reversed a plan to sunset 4o, an earlier GPT model known for its sycophantic conversation style, amid intense backlash. But by February 2026, 4o was killed for good.

Even OpenAI can't escape the challenges

OpenAI CEO Sam Altman and his team have publicly delighted in Anthropic's struggles. OpenAI has continued to announce new tools and features at a faster clip.

Altman said he was sipping drinks on Tuesday night while Anthropic officials tried to tamp down confusion about Claude Code's availability.

"For clarity, we are running a small test for ~100% of Codex users where we: - make our best models available to all users - make Codex available to all plans, free and paid," Rohan Varma, who is part of OpenAI's Codex team, wrote on X. " Claude Code users aren't affected :)"

In a reply to Varma, Altman teased that OpenAI would soon roll out a rate limit reset for Codex, further enticing users to use OpenAI's coding tool.

In recent days, OpenAI has rolled out a new image generation model, ChatGPT Images 2.0, and cloud-based agents for some paid users.

But OpenAI has shown its willingness to make tough decisions, too. Last month, OpenAI announced it was ending Sora, its once-popular TikTok-esque AI video-generation app.

"OpenAI certainly has a compute moat through its deep integration with Microsoft's Azure infrastructure and its own massively secured compute," Chandrasekaran said, "but that doesn't mean they are immune to the pricing model challenges and the way AI unit economics work."