Theo says he confirmed a bug on Anthropic's side that makes Claude Agents SDK usage drain five hour limits faster than Claude Code
Spent a lot of time on this today. I confirmed it's a real bug on Anthropic's side. To be clear: this does not affect weekly limits. All Agents SDK usage will drain your 5h limits faster than using Claude Code directly. I have reported this to Anthropic, hoping for a fix soon.
Context
Many third party coding tools run Claude through the Agents SDK on users' own subscriptions, so a metering bug quietly cuts how much work those users get. The finding is from a developer's own tests, and Anthropic has not publicly confirmed it.
Theo Browne runs T3 Code, a coding app that uses Claude through Anthropic's Agents SDK. The post quotes a user complaint that limits were vanishing, and an open issue in Anthropic's Claude Code repository reports similar results. The post was published at 9:25 PM Eastern on October 24.
Deep dive
Theo posted at 9:25 PM Eastern on October 24 that he spent much of the day on the issue and confirmed it is a real bug on Anthropic's side. He said all Agents SDK usage drains the five hour limit faster than using Claude Code directly, that weekly limits are not affected, and that he has reported it to Anthropic.
The screenshot in his post shows the test behind that claim. For the same tokens, T3 burned the five hour limit about 1.7 times faster than the interactive Claude Code interface, while the weekly limit moved at the same rate. The summary attributes the gap to how Anthropic meters Claude Code's non interactive modes, which T3 uses through the SDK, and says T3 was not sending extra tokens or breaking the cache. In the table, interactive runs added about 7% to the five hour meter while T3 and headless runs added between 9% and 16%, with every run adding 2% to the weekly meter.
An open issue in Anthropic's Claude Code repository reports a similar result, with headless usage taking about 1.8 times more of the five hour window per token than the interactive interface. The quoted post came from a user who said his limits kept disappearing.
Anthropic had not publicly acknowledged the bug when this deck was compiled, so the diagnosis rests on Theo's tests and user reports.
What to expect next
Watch for an Anthropic response on the GitHub issue or from its developer accounts, and whether a fix restores the expected rate or comes with any usage credit for affected subscribers. Until then, developers building on the Agents SDK may see users hit limits early and blame their apps. A confirmed fix would also matter for Anthropic's push to keep developers building on its own SDK rather than rival models.