Anthropic’s AI Threat Hype Crushed by a Harsh Shutdown of Claude Fable 5

Anthropic repeatedly warned that AI could be dangerous, promoted its own models as the most powerful, and then released Claude Fable 5 (public version with safety guardrails) and Claude Mythos 5 (unrestricted elite version) on June 9, claiming they were the strongest models available to the public.

Timeline of Events

Early April : Anthropic released the Claude Mythos preview (Project Glasswing), stating its ability to discover software vulnerabilities was too strong for open release and limited it to about 50 security‑focused partners such as Amazon, Apple, Google, Microsoft, and CrowdStrike.

April 21 : Sam Altman described the approach as “fear‑based marketing” on the Core Memory podcast.

June 9 : Official launch of Claude Fable 5 (public with guardrails) and Claude Mythos 5 (elite without guardrails), advertised as the current strongest publicly usable models.

June 12 : Anthropic announced a global shutdown of both models for all users, while other models remained unaffected.

Model Relationship and Guardrails

Both Mythos 5 and Fable 5 are built on the same underlying Mythos‑level model; the difference is that Fable 5 adds several safety classifiers (guardrails) that filter outputs, whereas Mythos 5 has none.

Key Differences

Underlying model : identical for both.

Safety classifiers : Mythos 5 – none; Fable 5 – full set of guardrails.

Availability : Mythos 5 limited to trusted partners in Project Glasswing; Fable 5 released to the public.

Guardrail hit rate : Less than 5 % of requests are blocked and fall back to Opus 4.8.

Data retention : Mandatory 30‑day retention for both.

Price : 50 USD per million output tokens (same for both).

Safety Guardrail Impact

Anthropic’s own interception curve for network‑security tasks shows the guardrails in Fable 5 suppress attack progress almost entirely, effectively “killing” the attack on the floor.

Benchmark Highlights

Stripe migration : Fable 5 migrated a 50‑million‑line Ruby codebase that would have taken a team two months to do manually in just one day.

Cognition FrontierCode evaluation : Fable 5 achieved the highest score among all frontier models, even under moderate reasoning intensity.

Visual task : Previously, playing Pokémon FireRed required extensive map tools; Fable 5 solved it by merely looking at screenshots.

Long‑context task : In the card game Slay the Spire, providing a file‑based memory system let Fable 5 perform three times better than Opus 4.8.

Drug design : Mythos 5 identified strong candidates for 9 out of 14 protein targets; an independent E. coli experiment later validated the results.

Genomics : Mythos 5 autonomously processed over a week of data from millions of cells across 138 species, producing a small model that is 1/100 the size of a recent Science paper’s model yet delivers better performance.

Speculation on the Shutdown Motive

Earlier discussions questioned whether Anthropic’s restriction of Mythos 5 was meant to protect the internet or Anthropic itself. The author argues the likely motive is commercial: preventing other companies from distilling the model while deepening enterprise‑level contract barriers.

Community Reaction

Codex’s response to the situation was summed up as “That’s a good thing.”