One model, two names
Fable 5 and Mythos 5 share the same underlying weights. Fable ships with classifier safeguards active and is generally available. Mythos has those safeguards lifted in specific domains and is restricted to vetted partners. Anthropic's framing: Fable is "a Mythos-class model made safe for general use."
How Fable's safeguards work
Classifiers watch three areas. They trigger on less than 5% of sessions on average; when one trips, the response is handled by Claude Opus 4.8 instead — so 95%+ of sessions never see a fallback.
| Classifier | Blocks |
|---|---|
| Cybersecurity | Vulnerability discovery and exploitation; offensive cyber tasks (reconnaissance, lateral movement) |
| Biology & chemistry | Bioweapons-adjacent queries — a wider net than previous models cast |
| Distillation | Requests flagged as attempts to extract the model to train competitors |
The red-teaming record is unusually concrete: an external bug bounty found no universal jailbreaks in 1,000+ hours; Fable complied with zero harmful single-turn requests across 30 public jailbreak techniques; and it held through 400-turn internal jailbreak attempts where other models gave way (UK AISI did report progress in a brief testing window). Charts: cyber evaluations · alignment assessment (Mythos's measured misalignment, including deception, was low and similar to Opus 4.8).
What Mythos did with the guardrails off
- Drug design — protein-design experts accelerated parts of their process ~10×. Mythos ran the full scientist's loop: choosing binding sites, running design tools, recovering from failures. 9 of 14 protein targets yielded strong candidates. The designed complexes.
- Molecular biology — scientists preferred Mythos hypotheses over Opus-class models' ~80% of the time in blinded comparisons; one novel E. coli mechanism hypothesis was corroborated by independent research.
- Genomics — a week-long, largely autonomous run assembled single-cell data for millions of cells across 138 species and trained a custom model that beat a recent Science publication at 1/100th the size.
- Protein evals — outperformed dedicated protein language models on AAV property prediction using biological reasoning alone.
Who gets Mythos, and on what terms
- Project Glasswing partners — cyber safeguards lifted.
- Select biology researchers — bio/chem safeguards lifted; cyber stays on.
- A broader trusted-access program is planned for both domains.
All Mythos-class traffic carries a 30-day retention policy, is excluded from training and non-safety uses, and every human access to it is logged.
Moral: the same model can be a tool or a hazard — the difference is who's holding it, and what's watching.