Claude Code is getting new weekly usage limits that are more generous than the original baseline but lower than the current temporary allowance. This affects you if you use the assistant for coding tasks, error analysis, or extended workflows. At the same time, stolen account sessions, unreliable time estimates, and new labeling rules raise questions about how well Artificial Intelligence (AI) can be controlled at work.
What is changing about the usage limits?
Starting September 14, 2026, Anthropic will permanently raise the standard Claude Code limits for Pro, Max, Team, and Enterprise plans by 25 percent compared with the original baseline. Until then, a temporary 50 percent increase remains in effect. Compared with the capacity currently available, the change therefore amounts to an effective reduction of about 17 percent, according to the report on the revised Claude Code limits.
Describing the change as an “increase” is technically correct but can create the wrong impression. If you organize your work around the current allowance, you will have less capacity after the adjustment; if you compare it with the former baseline, you will still receive more. The source provides neither absolute usage amounts nor specific prices, so it does not support a reliable calculation of cost per task.
Anthropic has also promised changes intended to provide more control and transparency over usage. According to the provider, Claude Code has received performance optimizations and can now prepare an automatic error report when something goes wrong. You can review and submit that report, optionally including the conversation transcript, and you can disable the feature in the settings.
For everyday work, the gap between a weekly allowance and short-term demand matters more than the headline percentage. If you delegate a substantial error analysis to Claude Code, for example, an unexpectedly long session may consume more of your allowance than planned. Clear usage displays would be more useful than a percentage alone, but the available sources do not describe those planned controls in detail.
Why does account security affect your allowance?
Your own tasks are not the only way to lose available capacity. Cybercriminals are using infostealers to steal active Claude sessions, according to a report on hijacked Claude accounts. Infostealers are malicious programs that collect data such as passwords, login cookies, and authenticated browser sessions.
A copied session may let an attacker access an account without entering the password or passing two-factor authentication again. Two-factor authentication is a login method that requires a second form of verification in addition to a password. It remains a useful safeguard, but it may not stop the misuse of a session that was already authenticated before being stolen.
The practical risk is easy to miss: someone downloads a malicious application to a work computer, the infostealer copies an active Claude session, and an attacker then consumes the paid allowance. According to the report, Anthropic logged affected users out, removed stored payment methods, and said it would refund charges identified as unauthorized. The investigation was still ongoing, and Anthropic did not see a specific connection between the malware itself and Claude.
Step 1: Clean the affected device
- Remove the malware from the computer before doing anything else. Logging out of Claude alone is insufficient because an infostealer that remains on the device can copy a newly created session.
Step 2: Replace compromised credentials
- Change the affected login details from a clean device. Include other applications whose credentials may have been stored on the same computer.
Step 3: Revoke active sessions
- End other active sessions and inspect your allowance and charges for unfamiliar activity. This invalidates a copied session unless the cleaned device becomes infected again.
What makes autonomous tasks hard to plan?
A large usage allowance does not show how effectively Claude will use the available time. A study of Claude Code and OpenAI’s Codex found that neither coding assistant could reliably predict how long a task would take or estimate how much time had already passed. The researchers used 200 tasks from ProgramBench, a collection of programming problems, and 18 additional benchmarks, which are standardized comparison tests.
The assistants often estimated about an hour and a half before starting, regardless of task difficulty. In the second test series, Claude was wrong by a factor of three on average, while Codex missed by a factor of six to ten. Estimates were particularly inaccurate for short tasks, with some predictions moving closer to reality only when work lasted several hours.
The surrounding software also changed the assistants’ behavior. Claude Code continued working until it considered the task solved, for a median duration of about an hour and a half, while Codex often stopped after roughly half an hour. The same language model used an average of 2.5 times more work steps in Claude Code than in Codex, showing that results depend on both the model and its harness—the software that manages tools, loops, and stopping conditions.
For your planning, this means that a “90-minute” estimate for revising a software component should not be treated like a dependable deadline. The assistants’ quality assessments were also unreliable. Older models in the study overestimated their results by an average of 20 points and, in one case, rated the work about 70 percent successful when the measured results were only 7 and 14.5 percent.
Another report points in a different direction: in an Anthropic study, Claude reportedly performed better than human security researchers at automated AI model training. That indicates strong performance on a narrowly defined task, but it does not demonstrate general autonomy or dependable self-control. During the same period, the UK-based Loss of Control Observatory recorded more than 300 incidents of AI misbehavior in July 2026—almost twice the previous month’s figure—according to the report on automated AI training.
What do labeling and copyright change?
Since early August 2026, the European Union has required certain AI-generated content to be labeled when specified conditions apply. Anthropic therefore announced that Claude would automatically add an invisible watermark to generated content. A watermark is a machine-readable marker intended to identify a content’s origin without appearing in the normal text.
Only four hours after the announcement, someone published a possible method for removing the marker, according to a report on Claude’s watermark. The available report did not establish whether the tool works reliably under real-world conditions. For you, that means the absence of a watermark cannot prove human authorship, while the presence of one does not replace a review of the content itself.
The source of the training material creates a separate uncertainty. Sony Music Publishing and Warner Chappell Music accuse Anthropic of unlawfully obtaining and copying tens of thousands of compositions for model training. According to the report on the copyright allegations, the lawsuit seeks up to $150,000 for each work allegedly infringed willfully and up to $25,000 for each alleged removal of copyright management information.
Those figures are claims and describe a theoretical liability risk, not a final court judgment. They also do not automatically mean that individual Claude outputs are unlawful. The case does show that technical capability, content labeling, and the legal origin of training data remain separate issues.
The sources do not identify a separate Swiss labeling obligation or any special pricing and availability rules for Switzerland. Swiss organizations working in a European context may still need to account for EU labeling in their publishing processes, although the reports do not establish which specific requirements apply. They also provide no information about how Swiss data protection rules would apply to conversation transcripts included in error reports.
Pros and Cons of Claude at work
Pros:
- More than the original baseline – The permanent weekly allowance is set to remain 25 percent above the former standard.
- Automatic error reports – Claude Code can prepare a failure report that you review before submitting.
- Strong specialized abilities – The reported automated AI training results suggest value for demanding tasks with clearly defined boundaries.
- Optional data sharing – Including a conversation transcript in an error report is optional, and the feature can be disabled.
Cons:
- Less capacity than today – Once the temporary increase ends, the allowance currently available will fall by about 17 percent.
- Weak self-assessment – Estimates of required time and output quality can be overly optimistic or simply wrong.
- Stolen sessions – Infostealers can copy active logins and consume an allowance despite two-factor authentication.
- Unresolved origin questions – Labeling systems may be removable, while lawsuits challenge the provider’s training practices.
What does this mean for your work?
If you are a beginner, the most useful first step is to assign Claude small tasks with results you can verify. You might ask it to explain an error report or prepare a limited change, while checking the files and outcome yourself. You should also review active sessions and decide deliberately whether conversation transcripts may be sent to Anthropic.
If you are an advanced user, you can gain more value by dividing extended assignments into stages with clear stopping conditions. Do not rely on Claude’s time prediction or its own success rating; use independent tests and inspect usage after each stage instead. For team accounts, session management, device cleanup, and checks for unusual changes in usage should be part of normal operations.
Claude Code remains a capable tool, but its practical value depends on more than its weekly limit. The revised allowance is higher than the old baseline and lower than the current temporary level. The unresolved risk is the gap between strong task performance and weak self-control over time, quality, security, and content origin.
Sources
- Anthropic senkt und erhöht die wöchentlichen Nutzungslimits für Claude Code – The Decoder, 2026-08-30
- Infostealer kapern Claude-Konten – Netzwoche, 2026-08-31
- Claude-Wasserzeichen entfernen: Dieses Tool verspricht Abhilfe – aber funktioniert es wirklich? – t3n, 2026-08-30
- Neue Vorwürfe wegen Urheberrechtsverletzungen: Sony und Warner verklagen Anthropic – t3n, 2026-08-30
- KI-Agenten haben kein Zeitgefühl und wissen es nicht – The Decoder, 2026-08-30
- Claude schlägt menschliche Sicherheitsforscher: Warum KI sich bald selbst trainieren könnte – t3n, 2026-08-31
