Theo Browne on Claude Ultra, Five-Hour Limits and SDK Usage

Anthropic has renamed its subscription plans: the former $200 Max 20x plan is now Ultra, and verified Ultra accounts lose the five-hour limit. Weekly limits remain. I welcome the change, but verification and differences between Claude Code and apps using the Agent SDK deserve a closer look.

The old 5x and 20x labels referred to the five-hour allowance, not every limit. When Anthropic later introduced a weekly limit for Max 20x, it was ten times Pro's weekly limit, leading some people to assume incorrectly that it was twenty times. A product leader acknowledged the confusion and apologized; I think the revised labels are clearer.

At the time of this video, the plans are Free, Pro at $20 a month, Max at $100 and Ultra at $200. Each has different usage allowances. I find the top plan generous for ordinary development, although my own unusually heavy usage still stretches multiple accounts. Max continues to have a five-hour limit.

The two limits serve different purposes: a weekly allowance caps use over the week, and a five-hour window restricts how much can be consumed in a burst. My explanation is capacity management. Subscription usage that concentrates in a busy period competes for compute that could otherwise serve enterprise customers.

In T3 Code's weekday analytics, I see usage begin rising around 4 a.m. Pacific time and peak around 7–9 a.m., with a substantial evening lull. I observed about a twofold gap between low and high periods. A short window can spread consumption across those periods, and it can also reduce the chance that a runaway task burns a whole weekly allowance.

I connect Anthropic's earlier relaxation of Claude Code limits to its reported access to additional compute through a SpaceX arrangement. My point is that these windows manage GPU availability, not simply the nominal API price of each request. That is my interpretation of the capacity and business incentives.

The new removal applies to verified Ultra accounts. The announcement says most existing subscribers receive it first and newer accounts retain a temporary restriction until verification, with an email when it is removed. Anthropic says phone verification at checkout is separate; users can check the usage settings for the current-session bar.

I dislike how little detail is provided about verification. Anthropic frames the temporary limit as fraud and abuse prevention; I speculate that large-scale collection of model outputs for distillation may also be relevant. That explanation is my inference, not a reason the announcement establishes.

There is another issue for T3 Code users. A public report says headless Claude Code or SDK-style use consumes roughly 1.8 times as much of a five-hour window per token as the interactive CLI for the same work. I checked my own calculations and say they agree, although this is an observed discrepancy rather than an independent controlled study.

Removing the window for verified Ultra users helps a significant share of our users, but Pro and Max users can still hit the shorter limit. I say I have raised the discrepancy with Anthropic and was frustrated by the lack of a fix. The same concern matters to third-party applications that integrate via the Agent SDK.

During recording I explain that T3 Code reports an SDK entry point and discuss changing that field to identify the actual application. Then I receive what I describe as tentative permission to merge a change. I expect it to align the five-hour usage treatment with the interactive CLI, but the outcome is prospective in the video rather than a demonstrated guarantee.

I also describe my own use of a CLI proxy and multiple accounts, while explicitly saying I cannot guarantee that setup is permitted or that accounts will not be banned. My main point is that the plan rename is modest, while the treatment of short-window limits and integrations has practical consequences for developers.