Claude Fable 5 and Claude Mythos 5 launch – state-of-the-art Mythos-class model with safety safeguards and tiered access

TL;DR

Anthropic announced the general‑availability of Claude Fable 5, a Mythos‑class model that outperforms all previously released Anthropic models on a wide range of benchmarks, and the limited‑release Mythos 5, which provides the same capabilities with fewer safety restrictions for vetted cyber‑defenders and, later, biology researchers. Both models are priced at $10 / M input tokens and $50 / M output tokens.

What is Claude Fable 5?

Claude Fable 5 is a Mythos‑class 1 model that Anthropic has deemed safe enough for unrestricted public use. It delivers state‑of‑the‑art results on nearly every AI capability benchmark tested, with especially large margins on long‑horizon, complex tasks such as large‑scale code migrations, multi‑step scientific reasoning, and vision‑only interactions.

Core capabilities

Software engineering – In a Stripe test, Fable 5 completed a 50‑million‑line Ruby codebase migration in a single day, a task that would normally require two months of manual work. On Cognition’s FrontierCode benchmark, it achieved the highest score among frontier models while using fewer tokens.
Knowledge work – On Hebbia’s Finance Benchmark, Fable 5 attained the top score for senior‑level reasoning, excelling at document‑based analysis, chart interpretation, and expected‑value calculations. IMC reported that the model “aced” their trading‑analysis suite across factual lookup, root‑cause analysis, and valuation tasks.
Vision – Fable 5 can extract precise numeric data from scientific figures, reconstruct web‑app source code from screenshots, and play Pokémon FireRed using only raw game images, surpassing earlier Claude models that required elaborate helper harnesses.
Long‑context memory – The model remains coherent over millions of tokens and can improve its output by taking persistent notes. In the deck‑building game Slay the Spire, a file‑based memory extension tripled its success rate compared with Claude Opus 4.8.
Life‑science research – Using the Mythos‑class core, Fable 5 (via Mythos 5) accelerated protein‑design pipelines by ~10×, generated novel molecular‑biology hypotheses that were preferred 80 % of the time over Opus‑class outputs, and produced a genomics model that outperformed a recent Science publication despite being 100× smaller.

What is Claude Mythos 5?

Claude Mythos 5 shares the exact underlying model as Fable 5 but relaxes several safety classifiers. It is currently deployed through Project Glasswing in partnership with the U.S. government and is available to a small cohort of cyber‑defenders and critical‑infrastructure providers. Future expansions will include a trusted‑access program for biology researchers.

Distinctive aspects

Cybersecurity focus – Mythos 5 possesses the strongest offensive and defensive cyber capabilities of any publicly known model, enabling rapid vulnerability discovery, exploit generation, and multi‑stage attack planning.
Reduced safeguards – The model bypasses the default fallback to Claude Opus 4.8 for cybersecurity, biology, chemistry, and distillation queries, allowing full use of its capabilities in trusted contexts.
Pricing parity – Mythos 5 is priced identically to Fable 5 ($10 / M input, $50 / M output), representing a cost reduction of more than 50 % compared with the earlier Claude Mythos Preview.

Safety architecture and safeguards

Anthropic introduced a conservative safety stack for Fable 5 to mitigate misuse while keeping the model broadly usable.

Classifier‑based request routing

Scope – Three classifier domains: cybersecurity, biology/chemistry, and model‑distillation.
Mechanism – When a classifier flags a request, the system automatically routes the query to Claude Opus 4.8, a highly capable fallback model, rather than refusing outright.
Impact – Fallbacks occur in less than 5 % of sessions on average; over 95 % of sessions proceed without any redirection.

Robustness against jailbreaks

Internal red‑team tests and an external bug‑bounty (1,000+ hours) found no universal jailbreaks.
An external partner reported zero successful malicious cyber queries across 30 different public jailbreak techniques.
Comparative graphs (see source images) show Fable 5’s safeguards resisting jailbreak attempts more effectively than prior generally‑available models such as Opus 4.8.

Data retention policy

All traffic on Mythos‑class models is retained for 30 days only for safety‑related analysis; it is not used for training new Claude models.
Anthropic logs all human access to retained data and deletes it after the retention window, providing additional privacy safeguards.

Early customer feedback

Multiple early‑access partners reported dramatic productivity gains:

CursorBench – Fable 5 set a new state‑of‑the‑art record, unlocking long‑horizon problem solving.
GitHub – Developers observed autonomous, reliable handling of complex, multi‑step coding tasks, enabling a shift toward “hand‑off” agentic workflows.
Legal review – Lawyers noted that Fable 5’s contract red‑lining matched or exceeded existing models.
Finance & analytics – The model broke the 90 % threshold on Anthropic’s core analytics benchmark, a 10‑point jump over Opus 4.8.
Scientific research – Fable 5 achieved near‑full coverage on a physics research benchmark in 36 hours, a speedup comparable to four days of GPT‑5.5.

(Full quotations are reproduced as image captions in the source; the above bullet points synthesize the reported sentiments.)

Pricing and availability

Pricing – $10 per million input tokens, $50 per million output tokens (less than half the cost of Claude Mythos Preview).
General access – Claude Fable 5 is available to all API users immediately.
Tiered rollout for subscription plans – From launch through June 22, Fable 5 is included at no extra cost on Pro, Max, Team, and seat‑based Enterprise plans; after that date usage credits are required until capacity permits broader inclusion.
Mythos 5 – Currently limited to Glasswing partners; a trusted‑access program for biology researchers will open later.

Implications for the AI ecosystem

Performance frontier – Fable 5 establishes a new benchmark for autonomous, long‑context reasoning, narrowing the gap between AI and human senior‑level experts in software, finance, and scientific domains.
Safety‑first deployment – By routing high‑risk queries to a slightly less capable model, Anthropic demonstrates a pragmatic approach to releasing powerful models while containing misuse potential.
Economic impact – The sub‑half‑price offering may accelerate adoption of frontier models in enterprise workflows, potentially reshaping the cost structure of AI‑augmented development and research.
Policy relevance – The tiered trusted‑access program and 30‑day data‑retention rule illustrate a concrete model for balancing openness with national‑security concerns, a template other frontier labs may emulate.

All figures, benchmark tables, and illustrative images referenced above are reproduced from Anthropic’s official announcement and system card.