Detailed Analysis
A Reddit user reporting in Portuguese describes an apparent discrepancy in the usage counter behavior of Anthropic's Claude MAX 20x subscription plan. According to the post, the user's weekly Sonnet model usage was sitting at approximately 70%, with the aggregate across all models also near that threshold. Upon returning from a brief absence — described as a coffee break — the user found that the total usage counter had jumped to 100% without any corresponding increase in the Sonnet-specific usage metric. The user accompanied the report with a screenshot link as evidence of the inconsistency.
The core of the user's argument rests on a logical constraint they had deliberately built into their workflow. Their system is configured with an environment variable that forces all API or interface calls to route through Claude Sonnet, explicitly preventing accidental usage of the more resource-intensive Claude Opus model. Under this setup, the user reasons it would be mathematically impossible to exhaust the total usage pool — which encompasses all models — without first exhausting or coming significantly closer to the Sonnet-specific weekly limit. The sudden 30-percentage-point jump in total usage, absent a proportional change in Sonnet consumption, is what the user identifies as the bug.
This report highlights a broader challenge Anthropic faces as it scales tiered subscription products like the MAX plan: the complexity of usage accounting across multiple models, rate-limit windows, and aggregation logic creates numerous potential failure points in what users experience as a transparent metering system. When limits are opaque or behave unexpectedly, even technically sophisticated users — as evidenced by this individual's use of environment variables to control model routing — are left without the means to audit or trust the counter. The absence of a granular usage log or activity breakdown makes it difficult to distinguish a genuine billing bug from a legitimate but non-obvious consumption event, such as background API calls or session overhead.
Within the broader context of AI subscription models, this incident reflects a growing tension between the complexity of consumption-based pricing and user expectations of predictable, auditable usage. As Anthropic and competitors like OpenAI expand tiered plans with differentiated model access, the accuracy and transparency of usage meters becomes a key element of user trust and product integrity. A counter that jumps unexpectedly — especially by a significant margin and without user-initiated action — undermines confidence in the fairness of the metering system and raises questions about whether users are being charged or throttled for consumption they did not knowingly incur. Anthropic has not publicly addressed this specific report, and without an official response or granular usage data, the root cause remains ambiguous.
Read original article →