Anthropic released Fable 5 and secret Mythos 5 without safety filters — via Project Glasswing

Fable 5 — flagship model of the new Mythos class — appeared a week after Dario Amodei's latest interview, in which he called on the industry to slow down AI development. At the same time, Anthropic announced Mythos 5: the same architecture without a safety classifier, available only by invitation through Project Glasswing. The scheme turned out to be more well-thought-out than it appears at first glance.

Fable 5 and the Mythos Class

Fable 5 opens a new line at Anthropic, positioned above existing Claude 4.X models. Official benchmarks demonstrate significant gains:

• SWE-Bench Pro: 80.3 points versus 69.2 for Opus 4.8 — a gap of over 11 points
• FrontierCode Diamond: results twice as high as its predecessor on the most complex engineering tasks
• Weight architecture is identical to Mythos 5 — both versions are built on the same foundation

For developers already using Opus 4.8 in agentic scenarios and complex coding, the numbers speak for themselves: tasks where the previous model made errors, Fable 5 solves confidently. At the same time, the Mythos class was not created to replace Sonnet or Haiku — those were optimized for cost and speed. Mythos pushes the ceiling of possibilities.

The release of Fable 5 also means that the gap between Anthropic's public models and what the company demonstrates in its internal systems has grown again. The next tier of the Mythos class has not yet been announced.

Mythos 5 and Project Glasswing

Along with Fable 5, Anthropic announced Mythos 5 — a version with the same weights, but without a safety classifier. Access is restricted: invitation-only through Project Glasswing. To appreciate the difference, it's important to understand what the standard filter does:

• Cybersec — blocks exploit instructions, infrastructure attacks, and malicious code generation
• Bioweapons — filters requests related to synthesis of dangerous pathogens and chemical substances
• Model distillation — rejects request patterns aimed at copying model architecture by third parties

Mythos 5 has none of these restrictions. The intended audience is security researchers, red teams, corporate partners who need full access for legitimate tasks: testing defenses, developing countermeasures against real threats, vulnerability research.

The closure through Glasswing adds manageability: Anthropic itself controls who receives the unfiltered version.

The Logic of the Paradox

Mid-week of that period, Dario Amodei gave an interview with the thesis: AI is developing too rapidly, the industry needs to slow down. Seven days later — the release of the most powerful model in the company's history plus the announcement of a version without safety filters. The contradiction is obvious, but behind it lies a clear strategy.

Public rhetoric about caution is part of Anthropic's positioning as a "responsible lab." The safety classifier in the mass market product serves a broad market and reduces abuse risks. Mythos 5 through Glasswing is a separate tool for a trusted professional circle, for whom full access is exactly what they need for legitimate work.

The scheme with two identical weights and different filters allows Anthropic to occupy two positions simultaneously: leader in performance and a company concerned with safety.

A similar approach has already been used by other labs: enterprise-level access, different API configurations for different types of clients. Anthropic formalizes this into a named program with explicit participant selection.

What This Means

Fable 5 redefines the ceiling of possibilities in coding and technical tasks. Mythos 5 and Project Glasswing show how leading labs will manage access to the most powerful versions of their products: not through the open market, but through controlled programs with selection. An open question remains — the criteria for participation in Glasswing and how quickly the new performance ceiling will become the norm beyond the closed program.