On 9 June 2026 Anthropic shipped Claude Fable 5 alongside a restricted sibling, Claude Mythos 5 — the first two models in a new “Mythos-class” tier that sits above the Opus line. Read the announcement as a systems diagram, not a product launch, and the interesting part is not the benchmark column. It is the switch hidden behind it.

Here is the topology. Fable 5 and Mythos 5 are the same underlying model. Fable is the version Anthropic calls “safe for general use”: on flagged topics — cybersecurity, biology — your query is silently re-routed to the next model down, Claude Opus 4.8. Anthropic says the router triggers in under 5% of sessions and admits it sometimes catches harmless requests. Mythos 5 is the same engine with those safeguards lifted, handed only to vetted cyber-defenders through Project Glasswing, a US-government collaboration. One model, two masks, and a classifier deciding in real time which face you are talking to.

←TODAY: A single frontier model now answers as itself or as a quieter fallback depending on what a hidden classifier thinks your prompt is about. →3012: The studios that survive are the ones who drew their dependency graph while the switches were still visible. Fulcrum: Capability you cannot predict the source of is capability you cannot audit — and an un-auditable dependency is a single point of failure waiting for a bad morning.

The capability is real. On the agentic SWE-Bench Pro evaluation Fable 5 posts 80.3% against Opus 4.8’s 69.2%; on the harder FrontierCode Diamond set the gap widens to 29.3% versus 13.4%. The proof point Anthropic leads with: Stripe ran Fable 5 across a 50-million-line Ruby codebase and finished a repository-wide migration in a day that a team had scoped at two-plus months. As TechCrunch noted, this lands days after Anthropic publicly warned that AI was getting “too dangerous” — the company shipped its most powerful public model anyway, which is the tension worth holding onto.

For a working practice the load-bearing detail is not the score. It is the bill of materials. Pricing is $10 per million input tokens and $50 per million output — roughly double Opus 4.8, the most expensive major model on the desk. Distribution arrived day-one on Amazon Bedrock and GitHub Copilot, but the GitHub rollout ships off by default and requires data retention enabled so Anthropic can run its safety classifiers. That is the sharp trade-off, stated plainly: to use the safe model inside Copilot, you let your prompts persist for inspection. For an office handling competition entries or client BIM data under NDA, that is a confidentiality decision, not a settings toggle.

Atelier: The Stripe migration maps directly onto the legacy work that clogs every computational practice — a decade of GDL macros, an IFC export pipeline nobody dares refactor, a Grasshopper definition held together by one person’s memory. This is exactly the long-horizon, too-big-for-one-engineer job PAZ built PAZGPT to sit on top of, and Fable 5 is the kind of engine that makes “talk to your Archicad model and have it rewrite the toolchain” a same-week task rather than a someday one.

Hack: This Hack teaches you to find the hidden single points of failure in your own AI-assisted toolchain before one of them finds you. The medium is a five-line Python move using networkx: model your stack as a graph where an edge means “A breaks if B is down,” then ask for the articulation points — the nodes whose removal splits the workflow in two. Run it and read the list out loud; the third name on it is usually the one you forgot you depended on.

import networkx as nx

g = nx.Graph()
g.add_edges_from([
    ("Archicad", "PAZGPT"), ("PAZGPT", "Fable5"),
    ("Grasshopper", "Fable5"), ("IFC_export", "Archicad"),
    ("Fable5", "Opus48_fallback"),  # the silent switch is a real edge
])
spofs = list(nx.articulation_points(g))
print("single points of failure:", spofs)

The EU AI Act treats a model at this tier as general-purpose AI “with systemic risk” (Article 55), pulling adversarial testing and incident reporting onto whoever deploys it. Switzerland sits outside the Act but downstream of it: a Zurich studio serving German clients through AWS inherits the obligation anyway. Don’t wait for that to become a fire drill. Draw your real dependency graph this week — not the architecture diagram, the dependency graph — and put the AI vendor on it as a node like any other.

Source: PAZ Inbox (human curated)