Anthropic’s Claude Fable 5 and Mythos 5: Twin Models with a Shockingly Low Price and New Safety Switches

Overview: Twin Models, Different Safety Settings

Anthropic launched two new models, Claude Fable 5 and Claude Mythos 5 , which are built on the same underlying architecture. The only distinction is the "classifier switch": Fable 5 is fully open (all safety classifiers enabled), while Mythos 5 disables the network‑security layer, allowing unrestricted access to certain topics.

Both models provide 1 M tokens of context and a maximum output of 128 k tokens, priced at 50 per MTok, a reduction of more than half compared with the previous Claude Mythos Preview (125 per MTok).

Key Highlights

The two models share the same base model; the difference lies in the safety‑classifier toggle.

Price: 50 per MTok, roughly a 60% cut from the prior limited model.

Context window of 1 M tokens puts them on par with flagship GPT series; agentic coding is highlighted as a strong use case.

Performance Stories

Story 1 – Stripe Ruby codebase migration : Stripe used Fable 5 to migrate a 50 million‑line Ruby codebase. What would have taken a small team two months was completed in a single day.

Story 2 – Cognition FrontierCode benchmark : The FrontierCode benchmark, created by 36 open‑source project maintainers, required 40 hours of scoring criteria per task. Fable 5 achieved the highest score in the “medium effort” tier and was the first model praised for acceptable code style.

“FrontierCode is scored by tech‑lead judgment, not CI,” – Celery author Tomer Nosrati.

Story 3 – Pokémon FireRed visual completion : Earlier Claude versions needed extensive map‑state tooling to progress. Fable 5, using only raw screen pixels, completed the game from start to finish.

Story 4 – Slay the Spire long‑memory test : Providing a persistent notebook, Fable 5 improved performance by 3× and reached the final act three times more often than Opus 4.8.

Story 5 – Mythos 5 protein‑design acceleration : Internal Anthropic protein‑design experts ran Mythos 5 without human assistance on 14 therapeutic targets, generating nine drug‑candidate molecules. The experiment also produced a novel E. coli protein‑mechanism hypothesis later reproduced by an independent lab (biorxiv.org/10.64898/2026.03.12.711259v1).

Three‑Layer Safety Classifier (Fable 5)

Fable 5 incorporates three real‑time classifiers that automatically fallback to Claude Opus 4.8 when triggered:

Network‑security layer : Blocks any request involving reconnaissance, lateral movement, exploit, or evasion. External red‑team testing (over 1 000 hours) found no universal jailbreak.

Biochemical layer : Allows most biochemical queries but routes them to Opus 4.8 to prevent misuse such as designing dangerous viruses. The model can predict gene‑editing effects on AAV capsid assembly.

Distillation‑attack layer : Detects systematic attempts to extract model behavior (e.g., from DeepSeek, Moonshot, MiniMax). Over 16 million suspicious requests were intercepted.

Fallback occurs in less than 5% of sessions; in 95% of cases Fable 5 and Mythos 5 behave equivalently.

Project Glasswing Access

Mythos 5 is not publicly available via the standard API. It is distributed through Project Glasswing, a secure‑infrastructure program involving partners such as Cisco, AWS, Microsoft, CrowdStrike, JPMorgan Chase, Google Cloud, and the Apache Foundation.

Pricing and Release Schedule

API access for Fable 5 began on 2024‑06‑09 at 50 per MTok. The first two weeks (6/9 – 6/22) offered free inclusion in Pro/Max/Team/Enterprise plans; after 6/23 the model must be used with separate usage credits.

Data‑Retention Policy

All traffic to Mythos‑class models (including Fable 5) is now retained for 30 days. The data is logged for security purposes, not used for training, and is deleted in most cases after the retention period.

Suitability

Ideal for : Large‑scale codebase refactoring, long‑running tasks, data analysis, legal/financial research (reported 30% speed boost), and Vibe‑coding competitions (near‑full scores on ViBench).

Not suitable for : Red‑team security research (network‑security classifier may block legitimate work) and highly compliance‑sensitive enterprises (30‑day data retention).