The 5-Hour Mirage – Anthropic’s Diabolical “Moving Goalposts” Subscription

When you signed up for the Max 5x plan at $100 (€90+VAT), you didn’t just buy intelligence — you bought a static expectation of capacity. In the early days, “usage limits” felt like a solid wall: you knew where they were, and when you hit them, the rules of engagement were clear. Then Anthropic shifted the architecture of your access to a 5-hour session window with an undisclosed ceiling.

The Mechanics

Correction (2026-04-24): An earlier version of this article described the 5-hour limit as a true sliding window where individual tokens “drip out” continuously after aging 5 hours. That was wrong. The correct model is session-based and documented by Anthropic themselves.

Anthropic’s 5-hour limit is a session window, not a moving-tokens window. Here is how it actually works:

1. Your first message starts a 5-hour clock.
2. Every message you send during those 5 hours counts against the same session quota.
3. At session_start + 5h the entire counter hard-resets to zero — not gradually, all at once.
4. Your next message after the reset starts a fresh 5-hour session.

Anthropic’s own support page confirms this: “if you hit your limit at 2 PM, your next allocation begins at 7 PM, then 12 AM, and so on” (support.claude.com article 12429409). The authoritative reset timestamp is also surfaced to Claude Code’s statusline as a single rate_limits.five_hour.resets_at Unix epoch value (statusline docs) — if the window truly slid, no single reset timestamp would exist.

The word “rolling” in Anthropic’s copy means “cycles session-to-session”, not “continuously sliding”.

Why it still feels diabolical

The corrected model is simpler than the sliding one — but Anthropic’s actual tricks are about opacity, not mechanics:

• The Invisible Ceiling: Anthropic has publicly said it “adjusts” limits during peak hours (roughly 5 AM – 11 AM PT, per user reports). Because they never publish the exact token ceiling for your plan, they can shrink it during your workday without breaking their terms. Your “5x” multiplier stays — but 5× an unknown number is still unknown. Max 5x, Max 20x, Pro: no public cap figures for any of them.

• The Cache Tax & the 5-Minute Trap:

  Claude Code uses a prompt cache to avoid reprocessing the entire context on follow-up requests. The main agent’s cache lifetime is 1 hour; subagents only get 5 minutes. This means:

  If you pause for more than 5 minutes (subagents) or more than 1 hour (main agent), your next request triggers a full cache rebuild.
  With large contexts (up to 1M tokens), a single cache miss can burn a significant chunk of your 5-hour session quota in one go.
  Tip: Before resuming a long, stale session, use /clear to deliberately reset the context rather than risking an expensive cache miss.
  Alternatively, set the environment variable CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 to cap context at 400K tokens.

  PS: Google’s Gemini does not have this trap.

• The Black Box: By refusing to publish token caps or ship a first-party visualisation tool, Anthropic forces you to guess when your 9 AM burst is going to hit the ceiling. That is exactly the gap Claude Token Vampire fills — it counts the tokens you are burning and shows you the real ceiling (the one you configure, since Anthropic won’t tell you theirs), plus a hard countdown to the session reset.

Update (and warning)

I discovered that if you activate “Extra Usage”, the numbers in that section will not be updated like the other ones when you press the small “Refresh” circle.

You need to refresh the whole page in the browser. Keep refreshing the page, otherwise you will get a nice surprise with many zeroes, like me 🙂

Sources

• Anthropic Support: Manage extra usage for paid Claude plans
• Anthropic Support: Usage limit best practices
• Anthropic Support: What is the Max plan?
• Claude Code: Customize your status line — documents rate_limits.five_hour.resets_at and used_percentage