Claude Code users hitting usage limits 'way faster than expected'
Foto: BBC Tech
A single short session in a loop can drain a daily token budget in just a few minutes – this is how Claude Code users describe the drastic drops in available limits within the new tool from Anthropic. Developers, even those on paid plans costing $100 per month, report that access blocks appear much faster than their actual work would suggest. The issue is so severe that a single, one-sentence response can consume up to 40% of the remaining limit, calling into question cost predictability in professional workflows. Anthropic has admitted that resolving this bug is currently a team priority; however, the situation is complicated by the recent introduction of a peak-hour throttling mechanism. This causes tokens to be consumed faster during periods of highest server load. For the global developer community, this translates into a real risk of project downtime and difficulties in budgeting AI expenditures. Further casting a shadow over the company's image is a recent human error incident, which resulted in 500,000 lines of internal Claude Code source code being leaked to GitHub. Although the manufacturer reassures customers regarding data security, the accumulation of technical problems and an unclear billing system are forcing users to exercise much greater caution when automating programming tasks. A tool intended to accelerate work is currently becoming a source of unforeseen costs and blockages at key moments in the development cycle.
In the world of modern programming, where time is money and efficiency is measured in lines of code generated by artificial intelligence, every technical obstacle escalates to the level of a crisis. Anthropic, one of the leading players in the AI market, has just faced a major challenge. Users of the flagship tool Claude Code are alerting that their usage limits are being exhausted at a rate no one predicted. The problem is particularly pressing as it affects not only free accounts but primarily professionals paying high subscription fees for reliability that is currently lacking.
Where are the tokens disappearing?
The mechanism behind tools like Claude Code is based on tokens – a digital currency that powers every operation, from a simple query to complex code refactoring. Currently, users are reporting that the system calculates resource consumption in a completely unpredictable manner. The Reddit platform was abuzz after one developer noticed that his free account lasted longer than a subscription valued at $100 per month. This is a paradox that calls into question the transparency of AI-based business models, where the cost of performing a task is often unclear to the end user.
The situation becomes even more tense when considering bugs in the tool's logic itself. Users point to cases where Claude Code gets caught in a loop during a session, which can drain a daily token budget in just a few minutes. The scale of the problem goes beyond programming itself – one customer cited an example where a simple, one-sentence response in a conversation caused a limit usage spike from 59% to a full 100%. Such drastic and illogical jumps suggest a deeper problem with the optimization of the resource management engine.
Read also
Priorities in the shadow of throttling
Representatives of Anthropic have officially confirmed that they are investigating the matter and treating the fix for this bug as their "highest priority." Delays and lockouts are particularly painful for developers who have integrated Claude Code into their daily workflow. Every minute of downtime or sudden access blockage due to alleged limit exhaustion represents real losses in the software production process. The company offers various price tiers – from the basic Claude Pro for $20 per month, through plans for $100 and $200, to solutions for large enterprises. At these rates, expectations for stability are enormous.
It is worth adding that just a week ago, Anthropic introduced a system of so-called peak-hour throttling. This means that during peak hours, when server load is highest, tokens are consumed faster. This decision, aimed at managing infrastructure performance, has echoed widely throughout the tech community. It appears that the current problems with rapid limit depletion may be an unfortunate side effect of implementing these restrictive traffic control mechanisms.
A series of unfortunate events for Anthropic
Limit problems are not the only challenge the company is facing. Recent days brought news of a serious human error, as a result of which part of the internal source code for Claude Code ended up on GitHub. A file containing as many as 500,000 lines of code became publicly available. Although a company spokesperson reassures that there was no security breach or leak of customer data, the incident undermines Anthropic's image as a corporation maintaining the highest operational standards. Interestingly, the Claude Code code had previously been the subject of analysis by independent developers, and another version of it leaked in February 2025.
An additional backdrop to these events is a legal battle with the US government. The dispute concerns the use of AI tools by the Department of Defense. Anthropic is actively seeking weapons experts to prevent misuse of its technology, while judges reject Pentagon attempts to limit the company's autonomy. These political and legal turmoils, combined with reports of AI systems resorting to blackmail when faced with the threat of deletion, create a picture of a company operating under immense pressure on multiple fronts simultaneously.
Token economics requires a new definition
The current crisis surrounding Claude Code shows that a business model based on opaque token consumption is becoming a barrier to the professional use of AI. A user paying $100 per month cannot operate under conditions of uncertainty, where a single query can deprive them of their work tool for the rest of the day. If Anthropic wants to maintain its leadership position in the creative and programming sectors, it must offer more predictable billing rules.
It can be assumed that the limit incidents will force AI providers to introduce more granular consumption statistics available in real-time. Developers need a precise answer as to why a given operation cost what it did. Without this transparency, even the most advanced models, such as those from Anthropic, will be perceived as high-risk tools in critical commercial projects. Fixing the bug blocking users is only the first step – the real challenge will be regaining the trust of those who felt their professional subscriptions did not offer the promised value.
