Skip to content
Underdog Diary

Nobody Knows Why Claude Bans Who It Bans

Accounts get banned with no violation anyone can point to. Other accounts break the rules openly for a year and nothing happens. Both are true at once.

  • Lesson 04
  • Beginner
  • 4 min
  • Updated: August 2026

Before this: None, but read it before you build anything serious on one account.

Current as of August 2026. Anthropic updates its Usage Policy and enforcement without much notice — see the DYOR note at the bottom.

There's no clean formula here, and that's the actual lesson. Accounts get banned with no violation anyone can point to. Other accounts break the rules openly for over a year and nothing happens. Both are true at the same time, and that combination is more useful to understand than any checklist of "don't do this."

The Numbers, Straight

  • July–December 2025: 1.45 million accounts banned, 52,000 appeals filed, 1,700 overturned (3.3%)
  • January–June 2026: 11.4 million banned, 398,000 appeals, 42,000 overturned (~10.5%)

That's almost an eightfold jump in banned accounts in a single half-year. Anthropic doesn't publish a reason breakdown — no category, no region split, nothing. So any post you find online claiming "X% of bans are for Y" is guessing. Treat those numbers as fiction.

Two Things That Are Both True

Bans land with no visible cause. Documented cases include a two-year user on a top-tier plan banned and refunded with no explanation given for which conversation triggered it, someone banned about an hour into their first session, and a 110-person company losing all access overnight — no warning, not even to the admins.

Real violations run for a long time untouched. One documented case: a person running thirteen parallel subscriptions to automate a retail business, left alone the whole time. The bigger one: routing third-party agent tools through a personal Claude subscription login worked openly for most of 2025. Anthropic didn't technically shut the door on it until January 9, 2026 — over a year after the workaround started spreading.

Put those two facts together and the conclusion is simple, if annoying: "it works for me" proves nothing. No ban today doesn't mean your usage is fine. It means enforcement hasn't reached you yet.

The One Confirmed Mechanic

Anthropic states directly that it uses your IP address to determine your region, and using the product from outside a supported region is a policy violation on its own — regardless of the reason you're doing it. That's the one piece that's officially confirmed. Anthropic doesn't publish further detail on how its detection works, and neither will this post.

If It Happens to You

  • Log in with the banned account itself. The appeal form shows up right there on the login screen — no separate site, no email needed to start.
  • There's no official response time. Documented cases range from one day to nine months. Plan accordingly.
  • Be direct, not emotional. One well-documented case was resolved in 24 hours with a single plain statement: account is personal, never shared with anyone. Stating the fact that undercuts the likely suspicion works better than arguing the ban itself was wrong.
  • The public Claude Code GitHub tracker isn't an appeal channel. Issues posted there get closed as duplicates with no company response. Use the actual appeal form.
  • Some people report a sync gap — a "your account has been restored" email arrives, but login still fails for a while after. It's been documented enough times to be worth knowing about rather than panicking over.
  • Refunds are issued in nearly every documented case. That part, at least, seems consistent.

Reducing Your Own Risk

  • Treat region restrictions as a policy line, not a technical problem to route around
  • Automate through the API and its own key — not by routing third-party tools through your personal subscription login
  • Don't spread one workflow across a pile of parallel personal accounts
  • If a business depends on Claude access, keep a fallback path — clean usage isn't a guarantee against an automated false positive
  • If you do appeal, lead with the fact that undercuts the likely suspicion, and expect it to take longer than a day

Sources

⚠️ DYOR: Enforcement policy and its consequences shift constantly and without much announcement. This is an August 2026 snapshot, not a permanent rulebook — check the links above before making a decision that matters to you.

New lessons, when they’re ready

No schedule, no drip campaign. I send one when I've actually learned something worth writing down.