I Cancelled Claude: Token Issues, Declining Quality, and Poor Support

· ai coding ai-agents · Source ↗

TLDR

  • A Claude Code Pro subscriber cancelled after three weeks of token limit confusion, cache double-billing, and copy-paste support responses that closed tickets without resolving them.

Key Takeaways

  • Pro plan’s five-hour token window depletes in two hours on a single project; a forced break resets the conversation cache, triggering a full codebase reload that costs tokens again.
  • An undocumented monthly usage limit warning appeared mid-session despite hourly and weekly limits showing green, with no explanation in Anthropic’s docs or settings page.
  • Weekly reset date shifted unexpectedly from the original day to Monday, with no communication from Anthropic.
  • Claude Opus proposed a generic JS initializer as a refactor shortcut instead of editing JSX directly; the workaround consumed roughly 50% of a five-hour session allowance before being corrected.
  • Support sent a docs-paste response about Pro/Max limits, closed the ticket with a no-reply notice, and never addressed the actual token spike after a ten-hour idle break.

Hacker News Comment Review

  • Commenters broadly agree that Claude Code Pro is borderline false advertising: technically usable but not practically workable at $20, while the $200 Max plan reportedly avoids the worst of the limits.
  • The capacity-constraint reading dominates: multiple commenters argue OpenAI, Google, and Anthropic are all quietly degrading products due to GPU supply pressure, letting price or throttling absorb demand rather than admitting it.
  • A thread-level concern about dependency risk ran parallel to the quality debate: building workflows on opaque, subscription-gated inference means the foundation can shift without notice, and open-source or self-hosted inference is increasingly competitive on quality.

Notable Comments

  • @drunken_thor: “AI services are only minorly incentivized to reduce token usage” – frames limit creep as a structural incentive problem, not a bug.
  • @zkmon: Found Claude Code agent consumed 10x more tokens than direct llama.cpp inference on the same task, pointing to agent orchestration overhead rather than model quality.
  • @petterroea: Argues Kimi 2.6 is near-Opus quality, suggesting Anthropic’s moat is brand, not capability, and commodity pricing may follow.

Original | Discuss on HN