Anthropic has unveiled Claude Fable 5, a new Mythos-class large language model that it is making generally available with additional safety controls, alongside Claude Mythos 5, a less restricted version aimed at vetted cybersecurity and life sciences users. The company positions Fable 5 as its most capable broadly accessible model to date, with performance gains across software engineering, knowledge work, vision tasks and scientific research, while emphasizing that the release comes with expanded protections against misuse.
Anthropic says Fable 5 surpasses its earlier Claude models on nearly all tested benchmarks, particularly on long, complex tasks. In early trials, partners used the model to compress months of engineering into days on a 50‑million‑line codebase and to handle demanding coding benchmarks at high quality standards. The model also posted strong results on finance evaluations requiring senior‑level reasoning and trading analysis, and set a new bar among Claude models for interpreting documents, charts and tables. In vision, Fable 5 is described as state of the art, able to reconstruct web applications from screenshots and to complete a full playthrough of the game Pokémon FireRed using only raw images, without additional tooling.
The underlying system for Fable 5 and Mythos 5 is the same, but Anthropic differentiates them through safeguards. Fable 5 routes certain queries in cybersecurity, biology and chemistry, and model distillation to the company’s next‑most‑capable model, Claude Opus 4.8, via new classifier-based safety systems. These classifiers are designed to detect attempts at misuse or jailbreaks and block or redirect responses, with Anthropic reporting that they trigger in fewer than 5% of sessions and that external red‑teaming has so far not produced a universal jailbreak. Mythos 5, by contrast, lifts these protections in some domains for a limited set of cyber defenders and infrastructure providers through Project Glasswing, and for selected life sciences researchers through a forthcoming trusted access program.
Anthropic frames the strengthened safeguards as a response to the dual‑use nature of frontier AI capabilities, particularly in offensive cybersecurity and advanced biological research. The company says it has already seen Mythos‑class models match or outperform specialized protein design systems in tasks relevant to drug development, autonomously generate molecular biology hypotheses that researchers have taken forward, and conduct genomics work that rivals a recent Science‑published model despite using a smaller architecture. To reduce the risk of uplift for malicious actors while still supporting beneficial research, Fable 5 currently falls back to Opus 4.8 on most biology and chemistry queries, while Mythos 5’s more powerful tools in those areas are restricted to vetted partners.
Alongside the new models, Anthropic is introducing a mandatory 30‑day data retention policy for Mythos‑class traffic, covering both first- and third‑party surfaces. The company says this data will not be used for training and will instead support safety monitoring, with access logging and post‑retention deletion commitments. Both Fable 5 and Mythos 5 are priced at 10 dollars per million input tokens and 50 dollars per million output tokens, less than half the price of the earlier Claude Mythos Preview. Fable 5 is available immediately via the Claude API and on Pro, Max, Team and seat‑based Enterprise plans at no extra cost through June 22, after which continued use on subscriptions will require usage credits until capacity allows it to be restored as an included feature. Mythos 5 remains limited to Project Glasswing partners and a small group of life sciences researchers until Anthropic expands its trusted access programs.
Read Article: Global Data Center Liquid Cooling Market to Reach $29.2 Billion by 2033
