Anthropic has a problem on its hands, and it’s not the kind you can patch overnight. According to reports shared on Reddit, users of Claude Code, the company’s AI-powered coding assistant, are burning through their token allowances at an alarming rate. Some are hitting limits they expected to last weeks in what feels like minutes.
The issue is getting enough attention that Anthropic has made fixing it a “top priority” for the team. But the real question here is whether this is a bug, a feature, or just the messy reality of how Technology companies price their services.
The Token Drain Nobody Saw Coming
Claude Code became popular because it actually works. Developers started using it as part of their daily workflow, and for the most part, it delivered. Then something changed. Or at least, that’s what users started noticing.
One Reddit commenter revealed they hit their token limit “much later” on their free account than they did on their $100-a-month paid tier. Another described sessions where “a simple one sentence reply to a conversation just took me from 59% usage to 100%.” That’s not a gradual resource drain. That’s a cliff.
The most damning comment came from someone describing how “one session in a loop can drain your daily budget in minutes.” If you’re a developer relying on Claude Code for your work, that’s not just frustrating. That’s a service disruption.
The Opacity Problem
Here’s where things get murky. Anthropic sells tokens like a currency. Customers load up their accounts with tokens and burn through them with each request. The problem is, the actual cost of any given task isn’t always obvious upfront. You might expect a simple code review to cost X tokens, only to watch it consume five times that amount.
This isn’t entirely Anthropic’s fault, necessarily. Language models are genuinely difficult to price accurately. Token consumption varies based on input length, output length, and the complexity of the reasoning involved. But from a user’s perspective, it feels like flying blind. You’re spending money, but you never quite know how much until it’s gone.
A Week of Bad Timing
Last week, Anthropic introduced peak-hour throttling across Claude, meaning tokens get consumed more quickly when demand is highest. The logic is sound from an infrastructure perspective. When the system is busy, charge more. But for users already complaining about unexpected costs, this announcement landed like a punch.
Claude Pro costs $20 a month. Higher tiers can run $100 or $200 monthly. For Business customers, custom pricing applies. These aren’t small amounts, and they’re predicated on predictability. If users can’t predict their costs, they’re essentially paying for a service they can’t fully rely on.
The Bigger Picture
Anthropic has been dealing with other fires too. The company recently leaked 500,000 lines of Claude Code’s internal source code on GitHub due to what a spokesperson called “human error, not a security breach.” No sensitive customer data was exposed, they say, but the incident added another layer of chaos to an already turbulent month for the company.
On top of that, Anthropic is tangled up in legal battles with the US government over how its tools can be used by the Department of Defense. When you’re fighting on multiple fronts, even a straightforward bug fix feels like it might fall through the cracks.
What This Actually Means
For developers, this matters because reliability matters. If you can’t predict your costs or trust that the service will stay within your budget, you start looking elsewhere. And Anthropic has plenty of competition right now.
The real lesson here is that pricing opacity kills adoption, no matter how good the product is underneath. Fixing the token consumption is necessary. But it won’t be enough if users still can’t see coming what they’re about to spend.
