Using Anthropic and OpenAI is like writing a blank check and depositing it into a black box. You’re authorizing spend you can’t see, against limits you can’t track, on work you can’t audit for efficiency. That’s the basis of a proposed class-action lawsuit by a Claude user who is suing Anthropic over its $100 Max 5x and $200 Max 20x coding tiers allegedly delivering far below the advertised capacity.
The plaintiff, Karl Kahn, says he blew through his caps almost immediately after upgrading, once burning 15% of his entire weekly allowance in a single five-hour session.
Here’s why the lawsuit has merit:
- There’s no live meter. When using the Claude app, nothing tells you “this response cost X tokens” or “you’re at Y% of your weekly cap”. You find out you’re at the end of the token road when you hit the brick wall.
- You can’t get estimates. At no point does Claude give you any indication of how many tokens you’re about to use before it begins burning them for a task.
- The caps are stacked, not simple. Usage runs against a five-hour rolling window AND a separate weekly cap at the same time. Anthropic employs two overlapping meters to track your usage, neither of which you can see.
- You can’t tell efficient work from wasted work. If a task takes three tries instead of one, Claude burns more tokens, but you can’t tell whether that’s because the task was genuinely hard or because the model did something wasteful.
- Claude fucks up all the time, and you still have to pay for it. Your tokens are burned whether or not Claude is successful at a particular task, and oftentimes it’s unsuccessful behind the scenes before producing a final output, which means you never even know about the wasted token usage.
The lawsuit targets whether Anthropic delivered upon its advertised usage limits. However, the real question is less about whether it’s delivering on all promised tokens, and more about whether it’s efficiently using those tokens. Anthropic is already moving away from unlimited and capped plans toward usage-based plans, so the “5x” or “20x” become irrelevant, but the question over how Claude uses your tokens remains.
When AI companies charge based on usage, how efficient they are with your tasks becomes a YOU problem, not a THEM problem. Even worse, there’s a structural incentive for usage-based models to be inefficient, as it drives up revenue.
I’m not claiming that AI companies are intentionally inefficient to drive up revenue, at least right now in 2026, as most are still subsidizing usage. However, at some point in the future, the question will change from “How efficient CAN we be?” to “How efficient MUST we be?” — especially when single-digit efficiency margins of error on token-usage, which may be indistinguishable to the user, equal billions of dollars a year to the company.
What are your thoughts?Join the conversation on my LinkedIn post.
