Subscriber demands refunds, alleging Claude Code headless mode drains Max plan limits ~1.8x faster
jyongchul · ghdev · 2026-11-04
A subscriber paying $1,400/month across seven Claude Max 20x plans filed a refund request in anthropics/claude-code (#107134), arguing that headless usage (claude -p, stream-json, entrypoint sdk-cli) consumes roughly 1.8x more of the 5-hour window per token than the interactive CLI for identical workloads.
Key evidence:
- Reproducible reports: issue #97074 (67 comments, no staff reply) confirms the 1.8x rate; related issues #95129 and #92305 report 15-16k fixed extra tokens per -p call and stdin input billed at 1.78x
- Official docs only state that SDK/-p usage draws from subscription limits, never disclosing a different rate; the user's accounts ran out of weekly limits days before reset
- Demands: a free month or full refund per affected subscription, plus an official statement on when and where the differential rate was disclosed
- Also references public claims (T3 Code author, Theo Browne video) that T3 Code was informally allowed to set the undocumented CLAUDECODEENTRYPOINT variable so its usage counts at the Claude Code rate
More from coding & agent
- Device MCP lets Claude Code and Codex control a real iPhone for UI testing — rudrank · 2026-11-04
- Dev builds melting-sandwich WebGPU animation with Gemini 4 Argon + Antigravity — AI_Andrew · 2026-11-04
- Sub-4B model agent benchmark: Qwen3.5-4B hits 51.4% strict pass, thinking mode adds up to 20 points — KienShen · 2026-11-04
- External deterministic state layer proposed to fix AI coding session amnesia, with 3 papers surveyed — ApprehensiveSkin7975 · 2026-11-04
- Axion 1.0: ByteDance-Backed Open Agent Foundation Model With 3-Stage Recipe — Yuqi Jia · 2026-11-04
- Caper Pre-Executes Shell Commands in Sandboxes for Safer Agent Approvals — SingaporeManagementUniversity · 2026-11-04