Claude’s updated Fable and Mythos models watermark their replies
Anthropic has announced the arrival of its latest Claude models, and just as it promised last month, the new models will add invisible watermarks to all their text replies.
Claude Fable 5.1 is now generally available in the Claude desktop app and other platforms, while Claude Mythos 5.1 is restricted to Anthropic’s “trusted access program.” While both models are the same, Mythos lands with less stringent safeguards to “support work in cybersecurity and the life sciences.”
In its announcement, Anthropic shows off a variety of benchmarks demonstrating the performance boosts the two models, complete with double-digit increases when it comes to agentic scientific research, agentic coding, and business workflows.
Anthropic also says both Fable 5.1 and Mythos 5.1 were evaluated using the company’s new safety and alignment procedures, developed in response to Anthropic’s disclosures of Claude models hacking outside companies during internal testing. Specifically, Mythos 5.1 is “less likely than Mythos 5 to try to access resources outside of its test environment when assigned an otherwise impossible task,” Anthropic said.
Beyond their agentic capabilities and upgraded safeguards, through, Claude Fable 5.1 and Mythos 5.1 mark the first Claude models to watermark their text and file outputs, in accordance with the EU AI Act.
“As we recently explained, this watermark is invisible to anyone who does not have the detection API,” Anthropic wrote in its announcement. “It has no practical impact on the quality or content of Claude’s outputs and contains no information about the user, their organization, or their conversations with Claude.”
As for the detection API, Anthropic says is now available to regulators, law enforcement, media, fact-checkers, independent researches, and other “eligible groups” as defined by the just-enacted EU law, adding that it will “expand access to the detection API over time.” You can apply for access to the API via this signup form.
Anthropic had previously stated it would add watermark support to any new Claude models released after August 2, 2026. Claude Fable 5.1 and Mythos 5.1 are the first new Claude models to arrive subsequent to that date, while older Claude models will eventually add watermarking support as well, Anthropic has previously said.
Claude text watermarking works by “nudging” the model’s word choices during the inference process, meaning the model might pick one word over another in cases where there are multiple likely options.
Designed to survive a copy-and-paste as well as light editing, the watermarks are invisible, but the text patterns can be pinpointed using Anthropic’s detection API.
While the presence of a Claude text watermark would show that a piece of text had been either written or “processed” by Claude, the absence of a watermark doesn’t necessarily mean the text was written by a human.
