Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
If you build on someone else's platform, their meter becomes your problem. A long-running bug report shows how that plays out.
Apps built on Anthropic's Agent SDK use up a subscriber's 5-hour Claude limit 1.6 to 2.9 times as fast as Claude Code's own apps for the same work, according to developers' controlled tests.
Claude plans cap use in 5-hour windows and by the week. The report has been open since September 25, and Anthropic has not replied in it. A developer still found the gap on November 12. Here is what the testers measured, the workaround they found, and why it may not be worth the risk.
Same Work, Faster Drain
The report's author is on Claude's Max plan. They had moved work from the terminal app to headless mode, which lets other programs run Claude Code. "We hit the 5h limit several times a day with no change in workload," they wrote, and they paid for extra usage that the same work in the terminal "would not have needed."
The 2.9x figure comes from one tester who compared Agent SDK and headless runs with runs tagged as the VS Code extension. "The same allowance lasts only about a third as long through them," they wrote.
Three Tests, Different Results
The measurements do not all agree, which is part of the story. The main results in the thread:
1.6x to 1.7x on the 5-hour meter. Theo, who runs T3 Code, an app on the SDK, tested it in October. "It's real. It's on Anthropic's side," he wrote, adding that he was "like 95% sure" a 5-hour boost Claude announced on September 22 "never got applied to Agent SDK traffic."
1.14x to 1.32x on the weekly limit. Theo had said "Weekly is not affected," but another tester later found Agent SDK requests used the weekly limit faster too.
No measurable weekly difference. A third tester saw no gap in weekly use on real work that was mostly cached and used tools.
Across these tests, the 5-hour gap was the most consistent finding, while the weekly results were mixed.
One Label on Every Request
Claude Code sends a label with every request saying where it came from: one value for apps on the TypeScript Agent SDK, another for headless runs. Testers found that setting an undocumented variable to a label name of their own brought usage down to about the terminal app's rate.
In one tester's comparison, measured as weighted tokens per 1% of the 5-hour window, headless runs with the default label got about 165K. The same runs with a custom label got about 357K, and the terminal app got about 285K.

Why the Fix Comes With a Warning
On October 28, a tester warned that a new label "doesn't fix the bug, it just hides that the requests come from headless 'claude -p'."
If the higher rate is intentional, they wrote, changing the label could count as getting around usage limits under Anthropic's Consumer Terms. "The real risk is getting your account banned."
Theo has since merged his own change labeling T3 Code's Claude requests "t3-code," and says his users are no longer affected. The thread does not record Anthropic's view on that approach.
When Users Blame the App
The usage gap sits on Anthropic's side, but complaints land on the app. When a user on X said on November 12 that T3 Code still "consumes 1.8x usage," Theo replied: "Fake news. Please be more specific, this is really damaging for us and it isn't true at all."
The next day, another user said they had left T3 Code for Claude Desktop over its usage. Theo told them it "hasn't been a thing for a few weeks now." Whatever the current numbers, the earlier gap showed up for the app as public complaints and at least one user who left.
What Builders Can Take From It
Most readers are not building coding agents. The lesson applies to anyone whose product runs on another company's plan or API:
Your users feel your vendor's limits. If a platform meters your customers, any change in that meter looks like your product getting worse.
Measure before you argue. The clearest evidence in this thread came from controlled comparisons, not impressions.
Read the terms before using a workaround. A fix that saves usage but risks an account ban moves the problem rather than solving it.
Keep a public record of what you tested. When complaints arrive late, dated results make it easier to show what changed and when.

