Mythos 5 to Glasswing, Fable 5 to the public — Anthropic operationalizes a two-tier safety regime

The simultaneous release of Mythos 5 (Glasswing-tier) and Claude Fable 5 (public-tier) makes Anthropic the first frontier lab to ship differential safety conditioning by customer access tier — and to disclose that's what it's doing.

The substantive structural shift this week is in Anthropic's twin release. Mythos 5 goes to Project Glasswing partners — JPMorgan, MUFG, SMBC, Mizuho, AWS, Apple, Cisco, Google, Microsoft, the Japanese Finance Ministry. Fable 5 goes to Pro/Max/Team/Enterprise subscribers. Both are Mythos-class architecture; the difference is the deployment-time safety layer.

The two-tier disclosure is the new pattern

Anthropic explicitly disclosed that Fable's public distribution is possible because of new safeguards that block responses in specific high-risk areas — meaning Mythos 5 doesn't have those blocks, because Glasswing partners are vetted and bound by access agreements that make the wider envelope safe. That's the first time a frontier lab has explicitly disclosed differential safety conditioning by customer tier.

The cybersecurity capability frames the access tier

Mythos's restricted April rollout cited advanced cybersecurity capability — the model could find vulnerabilities, write exploits, run sophisticated attack chains. That capability is the value proposition for the Glasswing tier: JPMorgan's red team wants the wider envelope; consumer Claude.ai users do not. Anthropic's positioning is essentially "the most capable model on the cybersecurity dimension, available only to organizations that can responsibly handle that capability."

The audit pipeline becomes procurement deliverable

Mythos 5 ships to Glasswing with interpretability audit documentation as part of the deployment package — not academic publications but contract-tier deliverables the partner's procurement team uses internally. That moves interpretability research from "publication artifact" to "product asset" inside Anthropic's enterprise stack. The methodology questions raised by DeepMind's SAE deprioritization become procurement-relevant immediately.

What this means for the broader regulatory frame

EU AI Act compliance treats the deployed system, not just the model weights. Fable 5 (with safeguards) and Mythos 5 (without) are arguably two systems for compliance purposes even though they share the same trained model. That's a clarifying signal for the August 2 transparency window and the GPAI Code of Practice: the labelling and disclosure obligations attach to the product, not the weights.

Anthropic — Claude Fable 5 and Claude Mythos 5 → · TechCrunch — Anthropic releases Claude Fable, a version of Mythos →