Microsoft, Meta, and AI startups are reducing their dependence on Anthropic's Claude. The problem isn't necessarily performance. It's the cost of staying.
Claude has been one of the biggest winners of the AI boom. Anthropic built a reputation for powerful models, and Claude Code became a favorite among developers and enterprises.
But something interesting is happening. Some of Claude's biggest customers are starting to pull back.
Microsoft has reportedly reduced internal Claude spending by a third. Meta has cut the number of employees using Claude Code in half. And AI startups are shifting workloads to cheaper competitors, including open-weight models.
The reason isn't that Claude suddenly stopped being good. It's that being good isn't enough when the bill keeps growing.
Even Microsoft and Meta Are Cutting Back
According to an October report from The Information, Microsoft has reduced internal Claude spending by approximately one-third as it encourages employees to use its own AI tools.
The company reportedly reduced its Claude Code users from approximately 60,000 to 30,000, while its internally developed alternative, MetaCode, surpassed 30,000 users. The Information also reported that Meta spent more than $105 million on Claude Code during one recent 28-day period.
These aren't companies that can't afford Claude. They're companies with enormous AI budgets, sophisticated engineering teams, and access to some of the most advanced technology in the world. And they're deciding that relying so heavily on an outside AI provider doesn't always make economic sense.
To be clear, neither company has abandoned Claude. Both remain significant users. But the reductions reveal how quickly AI purchasing decisions can change when companies have alternatives.
One Startup Moved Everything to DeepSeek
For smaller businesses, the economics can be even more consequential. Flo Crivello, CEO of AI startup Lindy, told CNBC that his company moved 100% of its Claude model traffic to DeepSeek. He expects the switch to save millions of dollars within months.
For a company whose product depends on executing large numbers of AI tasks, the difference between expensive and inexpensive inference isn't just an optimization. It can determine whether the business works financially. As Crivello explained, the decision was about survival.
Quartz's July report describes how enterprise customers are reassessing their spending on both Anthropic and OpenAI as costs rise.
And Lindy isn't alone.
The AI Price War Is Changing Customer Loyalty
In an October 8 report, The Wall Street Journal described growing competition between OpenAI and Anthropic as companies reconsider which models deserve their spending.
David Zhu, CEO of AI sales platform Reevo, said his company is shifting usage from Anthropic to OpenAI. His reasons were straightforward: cost and a desire to avoid depending on one model provider. The shift isn't necessarily permanent. Zhu acknowledged that his preference could change again as the technology evolves.
And that's precisely what makes this development important. For years, technology vendors benefited from customers building their businesses around proprietary platforms. Switching was difficult, expensive, and disruptive. AI customers are beginning to recognize that the best model for one task may not be the best model for another.
A sophisticated reasoning model might justify premium pricing for complex analysis. But using that same model for thousands of simple classification, extraction, or routing tasks may be unnecessarily expensive.
Companies are learning to match the model to the work. And increasingly, they're unwilling to pay premium prices for every task.
Open-Weight Models Are Changing the Equation
This isn't simply a competition between Claude and ChatGPT. Open-weight models are creating another option.
DeepSeek and other open-weight developers have made increasingly capable models available at significantly lower inference costs.
According to The Information, open-source models accounted for 65% of tokens processed through OpenRouter in June 2026. That measures activity on one platform, not the entire AI market, but it illustrates the scale of demand for alternatives.
Open-weight models also give organizations the option to deploy AI in environments they control. That can provide greater flexibility over infrastructure, data handling, and operating costs, although self-hosting introduces its own security and operational responsibilities.
The important point isn't that open models are always better or cheaper.
It's that companies have more choices. And those choices give them leverage.
This Doesn't Mean Claude Is Losing
It's worth separating a change in customer behavior from a prediction about Anthropic's future. Claude remains highly competitive, particularly in advanced coding and reasoning applications.
The Wall Street Journal reported that some companies continue to prefer Anthropic's most capable models because their work demands frontier-level performance. Anthropic also continues to attract substantial enterprise demand. But the company is competing in a market where technical leadership can change quickly, and customers are increasingly willing to switch providers when the economics favor it.
That creates a difficult challenge for every model company. They need to deliver enough differentiated value to justify their prices, even as competitors improve and lower-cost alternatives proliferate.
The best model doesn't automatically win. The best model for the job, at the right price, increasingly does.
The Real Story Isn't Companies Leaving Claude. It's Companies Leaving Lock-In.
Microsoft, Meta, Lindy, and Reevo are making different decisions. Some are developing internal alternatives. Others are switching providers. Still others are distributing workloads across multiple models.
But their decisions reflect a bigger shift. Powerful AI is no longer limited to a handful of proprietary models. Open-weight alternatives like DeepSeek are giving businesses and developers new options for performance, cost, and innovation.
And as more models become available, the next challenge is making them easier and more affordable to use. That's what we're working to change at Rival.
Our All-You-Can-Eat offering gives developers and businesses unlimited access to DeepSeek 4 Flash without subscription fees, per-token charges, or throttling. Users pay for the compute they consume and any optional paid tools, with the freedom to build agents, workflows, and applications without constantly counting tokens.
And DeepSeek is just the beginning. We plan to make additional open-weight models available, expanding what builders can access and create. The future of AI won't be defined by which model wins or which company loses ground. It will be defined by how many people can put powerful AI to work.
The Claude honeymoon may be ending. But a much more interesting era is beginning: one where access to powerful AI becomes easier, more affordable, and less restrictive for everyone.