Anthropic’s Claude Fable 5 and Mythos 5: 50 M‑Line Code Migration in One Day

Anthropic released two new Claude models—Fable 5, open to all users with a safety classifier, and Mythos 5, a restricted, high‑security version—both achieving record‑breaking performance on software‑engineering, research, vision, and long‑context tasks, while offering a pricing model of $10 per M input tokens and $50 per M output tokens.

SuanNi
SuanNi
SuanNi
Anthropic’s Claude Fable 5 and Mythos 5: 50 M‑Line Code Migration in One Day

Model variants

Claude Fable 5 and Claude Mythos 5 share the same underlying model. Fable 5 is publicly available and includes a safety classifier that intercepts requests in domains such as network security, biochemistry, and model distillation. When the classifier flags a request, the system falls back to Claude Opus 4.8 and notifies the user. Anthropic reports that >95 % of sessions have no fallback; fallbacks occur in <5 % of sessions. The classifier is intentionally conservative and may block harmless requests.

Mythos 5 removes several safety layers to expose the strongest network‑security capabilities. It is deployed via Project Glasswing in collaboration with the U.S. government and is initially limited to a small group of security‑focused customers. Anthropic plans a Trusted Access rollout, including a biomedical research track.

Benchmark performance

SWE‑Bench Pro (software‑engineering) – 80.3 % accuracy, +11.1 percentage points over Opus 4.8.

FrontierCode (Diamond) – +15.9 % over Opus 4.8; highest score among frontier models while using fewer tokens.

Stripe internal test – a 50 million‑line Ruby codebase migration completed in one day versus two months for a team.

Hebbia financial‑reasoning benchmark – top score across all models, large gains in fact‑finding, concept reasoning, root‑cause analysis, and expected‑value estimation.

IMC trading‑analysis suite – Fable 5 outperformed all competitors on the same dimensions.

Vision tasks – precise number extraction from scientific charts, source‑code reconstruction from screenshots, and full completion of Pokémon FireRed using only visual input.

Long‑context – maintains focus over million‑token sequences; in a Slay the Spire test with persistent file memory, performance and final‑chapter reach were three times higher than Opus 4.8.

Demonstrated capabilities

Autonomous tasks completed by Fable 5 include deriving planetary orbits from first‑principles physics, designing fluid‑simulation animations synchronized to classical music, and building complete 3D‑printed models in the browser with a self‑written CAD editor and AI copilot.

Research applications (Mythos 5)

Protein‑design pipeline accelerated ~10×; in a blind test of 14 targets, 9 yielded strong candidate designs matching or exceeding skilled human operators.

Novel molecular‑biology hypotheses preferred by ~80 % of scientists in blind comparisons; several entered experimental validation, including a newly proposed E. coli protein mechanism later confirmed by an independent lab.

Genomics study – integrated single‑cell data from 138 animal species (millions of cells), trained a custom ML model that outperformed a recent Science‑journal model while being only 1 % of its size. The study was completed in about one week with minimal human guidance.

Security and alignment

Mythos‑level models expose powerful network‑attack capabilities; Anthropic implements a three‑layer security classifier (network‑security, biochemical, and distillation) and a data‑retention policy that stores all Mythos traffic for 30 days for security analysis only, not for training.

Red‑team testing (>1,000 hours, external bounty program) found no universal jailbreak. External partners observed zero successful single‑turn network‑attack requests even when using 30 known jailbreak techniques. The UK AISI made limited progress but did not discover a universal jailbreak for long‑cycle agent tasks.

Biochemical requests are routed to Opus 4.8 pending a trusted‑access rollout for selected biomedical researchers.

Availability and pricing

Fable 5 is globally accessible via the Claude API under model name claude-fable-5. Pricing: $10 per million input tokens and $50 per million output tokens (approximately half the cost of the earlier Claude Mythos Preview). From the announcement date until 22 June, Fable 5 is included at no extra charge in Pro, Max, Team, and seat‑based Enterprise plans; after 23 June usage is billed against quotas.

Reference: https://www.anthropic.com/news/claude-fable-5-mythos-5

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

large language modelssoftware engineeringmodel safetyAI benchmarksresearch AIClaude Fable 5Mythos 5
SuanNi
Written by

SuanNi

A community for AI developers that aggregates large-model development services, models, and compute power.

0 followers
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.