FOLLOW US

Two new AI models just landed under one roof, sharing the exact same brain but wearing very different safety harnesses. Anthropic has released Claude Fable 5.1 and Mythos 5.1 — technically the same underlying model, distinguished only by the guardrails wrapped around each version, the company said.

Only One Version Reaches the Public

Access to these two models splits sharply down the middle. Fable is generally available, while Mythos remains restricted to vetted users through two dedicated access programs covering cybersecurity and life sciences. According to Anthropic, the biology-focused program was developed in collaboration with the U.S. government — a partnership that underscores just how sensitive the underlying capabilities are considered to be.

A Price Cut That Developers Will Notice

For developers, the headline news isn't the science — it's the price tag. Cache reads are dropping 75%, down to $0.25 per million tokens, a reduction Anthropic says translates into roughly 25% lower costs for typical workloads and up to 45% savings for agentic tasks — the kind of AI-driven, multi-step automated processes that tend to rack up token usage quickly.

A Benchmark Jump That's Hard to Ignore

The performance gains, at least by Anthropic's own reporting, are substantial. Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, more than double its predecessor Fable 5's score of 24.7% on the same benchmark. It also posted a 55.8% score on Terminal-Bench 4.0 — numbers that, if they hold up under independent scrutiny, would represent a significant leap in scientific reasoning capability within a single model generation.

The European Regulatory Thread

Buried further down in the announcement sits a detail with real regulatory weight: Anthropic signed the European Union's Code of Practice on transparency of AI-generated content back in July, joining roughly 190 other signatories in that commitment.

Why Claude's Output Now Carries a Watermark

That signature explains a quieter change baked into the release — Claude's output now carries a watermark. The mark applies specifically to models released after Aug. 2, 2026, remains invisible without a dedicated detection tool, and, according to Anthropic, contains no information tied to the individual user who generated the content.

A Detection Tool That's Narrower Than Expected

Reports last month suggested Anthropic would ship a free, publicly available detection API, letting anyone check whether a given piece of text carried the watermark. What's actually shipped falls short of that broader vision. The detection tool is currently in private preview, limited to eligible organizations — including regulators, law enforcement, media outlets, fact-checkers, independent researchers, educational institutions and EU civil society groups.

The Legal Backbone Behind the Watermark

The underlying obligation driving all of this is broad in scope. Article 50 of the EU's AI Act requires providers to mark synthetic content in a machine-readable format and ensure it remains detectable — a requirement that has been in effect since Aug. 2, 2026. Enterprises with their own separate compliance obligations can also apply for detection tool access, and Anthropic says it intends to expand availability over time, though it hasn't offered a specific timeline for doing so.

A Second Gap: Mythos Stays Out of Europe

There's a second notable European absence in this launch: Mythos itself. Its access programs, built specifically alongside the U.S. government, remain open only to a defined set of American organizations — a restriction that persists two years after Claude first became available in the EU.

Anthropic says it's working to eventually include international partners in these programs, but for now, European cyberdefenders and life scientists remain locked out of Mythos access entirely.

The Bigger Picture

Taken together, this launch tells a layered story: the watermark exists because European law demands it, the tool to actually verify that watermark remains stuck in preview, and separate reporting has already surfaced that tools capable of removing such watermarks are circulating as well — a reminder that regulatory intent and technical reality don't always move at the same pace.