Reallocating $100/Month Claude Code Spend to Zed and OpenRouter

Original Article ↗ Hacker News Discussion ↗

Claude Code pricing, limits, and quality

  • Several users report getting far more nominal “API value” from Claude subscriptions than they pay (e.g., hundreds or even tens of thousands of dollars of metered usage for $20–$200), implying heavy subsidy.
  • Others say new, tighter limits mean they now hit ceilings mid-session, especially for refactors and small web apps, making the $100 Max/Code tiers less attractive.
  • Some note noticeable quality degradation in recent weeks: more “lazy” behavior, missing edits, and poorer adherence to instructions compared to earlier this year.
  • Many still feel even $20 Pro is excellent value for planning, code review, and specs, but worry that individual-friendly pricing may not last.

OpenRouter fees, terms, and credit policies

  • The 5.5% fee (especially for BYOK after 1M requests) is seen as a meaningful cost, but many argue the unified API, routing, presets, multi-user keys, cost tracking, and zero-data-retention routing justify it.
  • OpenRouter states credits may expire after ~1 year of account inactivity; some see this as reasonable liability management, others as a potential “cash grab.”
  • Debate over what users are “allowed” to do: ToS prohibits reselling API access or building a competing API gateway, but allows building SaaS/products powered by OpenRouter.
  • Some confusion and frustration around “bans”: recent enforcement of upstream regional restrictions (e.g., OpenAI/Anthropic region rules) and crypto-related risk lead to blocked usage in some cases.

Editor & harness choices (Zed, VS Code, Pi, OpenCode, Codex)

  • Zed is praised for speed and Claude integration, but many report “papercuts”: weaker extension ecosystem vs VS Code, odd UX naming, Tailwind warnings by default, memory spikes, font/emoji issues, and missing workflows (e.g., hooks, some monorepo ergonomics).
  • Pi, OpenCode Zen/Go, and Codex are widely used as model-agnostic harnesses; Pi and similar tools are valued for transparency, extensibility, and keeping project knowledge in the repo instead of hidden memory.
  • Some dislike specific harness UX (e.g., OpenCode hover behavior) or reliability (reports of outages/slow speeds for some providers).

Model/provider mix and cost strategies

  • Common pattern: use top models (Claude Opus, GPT-5.x, etc.) for planning/review and cheaper models (GLM 5.x, Kimi, MiniMax, Qwen, etc.) for implementation to cut costs.
  • OpenCode Go, BlackBox, MiniMax plans, Z.ai, and Ollama Cloud are cited as extremely high-value subscription options compared to pure pay-per-token routing.
  • Others prefer GitHub Copilot or low-tier Claude/Codex subscriptions for predictable cost, but complain about per-request billing affecting workflow and aggressive context trimming.

Privacy, banking, and regulation

  • Some value OpenRouter’s ability to constrain providers to claimed zero-data-retention and to act as a privacy buffer between end user and model vendor.
  • A UK bank refusing to process OpenRouter payments (with unexplained reasoning) and crypto acceptance raise concerns about future regulatory and banking frictions around LLM intermediaries.

General sentiment on spend

  • Personal spend ranges from “$20 is plenty” to hundreds or even thousands per month, often justified as far cheaper than a junior developer.
  • Others are shocked at these amounts and see parts of the community as in “AI psychosis,” but heavy users report clear productivity gains and internal tooling built quickly enough to justify the cost.