Anthropic watermarks old Claude models and drops the 5-hour cap

Anthropic just put a mark on the text you already get from Claude. Retroactively. From November 3, 2026, the watermark also lands in the output of older model versions, not just the new ones. This isn't a one-month transparency pilot. It touches everyone who publishes text that came out of Claude.

What actually changes

• The watermark expands to previous Claude versions: from November 3, 2026 the mark goes there too (GIGAZINE).
• The paid tiers were renamed, prices untouched: Max 5x at $100 becomes plain Max, Max 20x at $200 becomes Ultra.
• Ultra removes the five-hour usage limit for verified accounts. The weekly cap stays (ITmedia).

Why Anthropic is doing it

Watermarking output answers the question regulators have asked for two years: how do you tell machine text from human text. While others draft codes of conduct, Anthropic bakes the mark into generation itself and asks nobody's permission. Doing it on older models is the message: transparency is default policy, not a pilot.

The pricing half looks smaller but follows the same logic. $200 for Ultra without the five-hour counter is an attempt to sell not tokens but the right to stop watching a counter. The weekly cap is still there. So the "unlimited" is formal, not real: the short ceiling is gone, the long one stays. Verified accounts are the filter against reselling access.

Where the mark breaks

A watermark in text is the weakest protection there is. It lives on generation statistics, not in a file. Retype a paragraph, translate it, paraphrase it in your own words — the mark is gone. Then the reverse problem: the finer the tuning, the higher the chance the mark fires on human writing.

Which leads somewhere uncomfortable. The mark catches the people who hid nothing: the honest author who worked in Claude. Anyone actually hiding a trail spends two minutes paraphrasing. The value isn't punishment, it's statistics: Anthropic gains a way to prove text came from them when that proof is useful to Anthropic.

What it means for the market

Only Anthropic marks. Rivals don't. So either the watermark becomes an industry standard, or it becomes a "made here" label — with the obvious downside: some clients walk to whoever stays silent. This is the old provenance fight, and Anthropic bet that transparency wins.

Prediction: Within a year, at least one major lab will announce a similar text-marking scheme, and independent detectors will publicly fight it over accuracy.

Disclaimer: not officially confirmed — based on Japanese media reporting, no direct Anthropic press release in our source registry.