Claude Code weekly rate limits

Hacker News Discussion ↗

Scope of the Change

  • Anthropic is adding weekly usage caps on Pro/Max plans, on top of the existing 5‑hour rolling window and monthly session caps.
  • They claim it will affect “less than 5% of users,” but it’s unclear if that’s 5% of all users, paid users, or Max users.
  • Example guidance: Max 20x “most users” can expect ~240–480 hours of Sonnet 4 and 24–40 hours of Opus 4 per week, but those are rough, non-binding ranges.

Fairness, “Abuse,” and All‑You‑Can‑Eat Analogies

  • Many compare this to an “all‑you‑can‑eat buffet” that quietly adds rules once people actually eat a lot.
  • One camp: heavy users are morally abusing a shared resource (24/7 agents, account sharing, running thousands of dollars of compute on a $200 plan) and “ruined it for everyone.”
  • Other camp: users simply used what was sold; the only “mistake” was Anthropic’s pricing and ToS. There’s no moral obligation to protect Anthropic’s margins.

Pricing Sustainability and Business Model

  • Widespread belief that flat‑fee access to frontier models is economically shaky given GPU and energy costs; estimates suggest sustainable plans would be in the high hundreds or thousands per month for true 24/7 heavy use.
  • Many see this as standard “price discovery”: loss‑leading to grab share, followed by tightening limits (“shrinkflation,” “bait and switch”).
  • Debate over future direction:
    • Some expect metered, per‑token billing to dominate.
    • Others predict more tiers (including very expensive ones) and eventual ad‑subsidized or “too‑cheap‑to‑meter” use for casual workloads.

Impact on Power Users vs Casuals

  • Some devs say they’re comfortably within limits; others report hitting Pro/Max caps quickly with “normal” coding (refactors, deep research, large repos), with no 24/7 agents.
  • Concern that the quoted 5% may mostly be the most serious, productive users rather than pure abusers.
  • Sub‑agents and “run lots of agents in parallel” were heavily promoted; now that same pattern is cited as a problem, which feels contradictory to many.

Opacity and UX Frustrations

  • Biggest practical complaint: no clear, official meter of usage or remaining quota; only vague “approaching limit” warnings.
  • Weekly windows are seen as especially punishing: one bad day or runaway process can lock you out for days.
  • Third‑party tools (e.g., ccusage) help, but users want built‑in dashboards, clearer numeric limits, and optional automatic fallback to pay‑per‑use API billing.

User Reactions and Alternatives

  • Some immediately canceled or plan to downgrade; others accept the change as necessary to keep the service viable and reduce outages.
  • A noticeable subset is looking at:
    • Gemini, ChatGPT/o3, Cursor, Roo Code, and other assistants.
    • Using Claude Code with their own API key (metered).
    • Building local or self‑hosted GPU rigs; trade‑offs between cost, power draw, and model quality are heavily debated.

Broader Reflections

  • Unease about becoming dependent on a proprietary, rate‑limited tool for daily work.
  • Mixed expectations: some think LLMs will eventually be cheap and ubiquitous like broadband; others fear “peak LLM” with persistently high prices and tightening limits.