Agents SDK may drain Claude's 5-hour limit faster than Claude Code

Checked: October 26, 2026. This is a time-sensitive note, not a product review. It summarizes a community-reported metering difference between Anthropic's Claude Code CLI/app and apps built on the Claude Agents SDK (for example T3 Code). We have not reproduced the bug ourselves. When Anthropic patches metering, this page should be demoted or archived. No affiliate links.

TL;DR

| Question | Short answer | |---|---| | What's wrong? | Same account, similar work: Agents SDK usage can empty the rolling 5-hour usage bar faster than Claude Code. | | Who said it's a bug? | Theo (X post, Oct 25) said he confirmed it on Anthropic's side, reported it, and noted weekly limits are not affected. | | What should you do now? | Prefer Claude Code directly for heavy sessions until a fix lands; keep SDK apps for light work. | | Weekly cap? | Reportedly unaffected. The pain is the 5-hour window. |

What people are reporting

On Oct 25, Theo wrote that he spent time confirming a real Anthropic-side bug: all Agents SDK usage drains the 5-hour limit faster than using Claude Code directly, weekly limits are fine, and the issue was reported upstream (post; ~2.4k likes / ~285 replies when we checked on Oct 26). Follow-up posts (for example KolinHuang and replies under Theo's thread) echo the same workaround: temporarily fall back to Claude Code.

Treat this as community confirmation, not an Anthropic status-page advisory. We have not seen an official Anthropic changelog entry that names this bug as of this check.

Practical workaround

1. Heavy coding / long agent runs: use Claude Code (terminal / desktop / IDE) on the same plan, so the 5-hour bar is metered the way you expect. 2. SDK-based apps (T3 Code and similar): keep for short tasks, or accept that the 5-hour bar may fall faster. 3. If you hit the wall mid-task: wait for the rolling 5-hour window to recover, or continue with API keys / usage credits if your plan allows (same advice as any other limit hit). 4. Do not assume the weekly cap is broken based on this report; reporters specifically say it is not.

How to keep evidence (optional)

If you want a before/after log for support, run one fixed prompt twice on the same account and note the 5-hour percentage (or remaining messages) before and after each run:

``bash # Example only: swap in your own fixed prompt file and tooling. # 1) Record usage % in Claude Code UI /status, then: claude -p "$(cat ./bench-prompt.txt)" # 2) Record usage % again. # 3) Repeat the same prompt through your Agents SDK app. # 4) Record usage % a third time. Save screenshots + timestamps (Asia/Shanghai). ``

We are not publishing a measured delta. Your numbers will depend on model, cache hits, and what the SDK app sends.

Related StackLoom pages

Live ranking entry: Top AI problems — Agents SDK 5h drain
Plan limits context: Claude Opus 5.5 vs Codex
Budget view: Which AI subscription to pick

Sources (checked 2026-10-26)

1. Theo — Agents SDK drains 5h faster; Anthropic-side; weekly unaffected: https://x.com/i/status/2114166439813017873 2. KolinHuang — summary / warning for T3 Code and other SDK apps: https://x.com/i/status/2114476596056776837 3. StackLoom Top AI problems entry for this week (Agents SDK 5h drain).