Theo reports Claude Agents SDK bug that drains five-hour session limits faster

• Agent SDK usage burns five-hour allowance faster than direct Claude Code, according to theo, who says weekly limits are unaffected.
• A controlled claude -p comparison found about 1.8× more quota consumption per token in the GitHub report theo shared.
• CLIProxy appears to avoid the discrepancy by adding a TUI header, theo reports, while he works with Anthropic on a fix.

An experiment on a Team account produced different quota consumption with the same SDK launch, binary and prompts, changing only the client’s entrypoint identity. In a separate reproduction, an environment variable inherited through tmux even made an interactive terminal session identify itself as SDK traffic.

Anthropic defines separate session limits that reset every five hours and weekly limits that reset once a week. Its Agent SDK policy says SDK calls, claude -p and third-party applications can draw from subscription allowances.

An October 7 update to that policy also introduced monthly API credits for Max and Team plans, while retaining subscription-limit access for SDK usage.

GitHub user qvarker opened issue #97074 on September 25 after testing Claude Code 2.1.282 on a Max subscription, using Opus 5 at high effort.

The workload and measurement setup were:

1. Fixed inputs: approximately 250,000 tokens across 18 files, followed by 41 prompts. The first prompt requested file reads; the remaining 40 asked short questions without tools.
2. Two execution paths: headless claude -p --input-format stream-json and the interactive TUI driven through tmux.
3. Normalized accounting: transcript usage deduplicated by message and request IDs, valued at API list prices, with other account activity measured and subtracted.

| Execution path | Test workload, API-equivalent value | Five-hour meter |
|---|---:|---:|
| Headless, sdk-cli | $12.74 | 52% → 60% |
| Interactive, cli | $12.91 | 60% → 65% |

Concurrent headless work contributed another $0.40 during the first phase and $1.12 during the second. After adjusting for that activity, qvarker estimated roughly 1.8× faster quota consumption headlessly.

The meter’s one-percentage-point resolution produced estimated ranges of $1.50–$1.90 of API-equivalent work per utilization point headlessly, versus $2.40–$3.90 interactively. These dollar amounts value transcript tokens at API prices.

qvarker also reported actual extra-usage charges after moving work from the TUI to headless mode, with logs showing rejected and isUsingOverage: true.

Anthropic’s programmatic-use documentation describes the Agent SDK as exposing Claude Code’s tools, agent loop and context management through the CLI, Python and TypeScript.

The sharpest follow-up came from GitHub user kler, whose October 17 experiment held the binary, prompts and environment constant while varying entrypoint labels on a Team account.

Relative to the claude-vscode-labelled control, the results were:

• sdk-ts, the TypeScript SDK’s default: 2.97× faster five-hour quota consumption.
• sdk-cli, the headless CLI label: 2.83× faster consumption.
• claude-vscode, an emulated extension identity: the reference rate.

All three arms ran through the SDK with headless streaming flags. The control emulated the VS Code extension by setting its entrypoint label; the experiment covered one account, plan and model, with a workload comprising about 96% cache reads.

The same issue includes a separate Pro-plan reproduction reporting approximately 2× faster headless consumption. Another contributor’s passive comparison against Claude Desktop found a smaller 1.3× gap across mixed workloads, with other-machine spending left unadjusted.

qvarker found fewer tokens of every type in the headless arm during a proxy-assisted rerun, with no retries or HTTP 429/529 responses. Model, effort, thinking and context-management settings matched.

The observed request differences were:

• cc_entrypoint=sdk-cli versus cli in the x-anthropic-billing-header system block.
• The User-Agent value.
• The redact-thinking-2026-02-12 beta, sent only by the interactive client.

Theo suspects Anthropic increased five-hour allowances without applying the boost to SDK traffic. His subsequent update identifies CLIProxy’s added TUI header as an apparent workaround and says he is working with Anthropic on a fix.

Boris Cherny, who works on Claude Code at Anthropic, asked affected users for recent session identifiers while the team investigated. He supplied a command that searches ~/.claude/projects for UUID-named session files modified within the previous three days and outputs deduplicated IDs.

The GitHub issue remained open as of October 25.