Anthropic unveils Fable 5 and Mythos 5 with hardened safety layers

Anthropic announced the launch of Claude Fable 5, its most capable model available to the public, alongside Claude Mythos 5, a restricted version limited to vetted cybersecurity and scientific research partners. The two models arrive with a suite of safety guardrails aimed at curbing misuse in fields such as cybersecurity, biology and chemistry.

Fable 5 outperforms every publicly released Claude model on a broad set of benchmarks, including software engineering, knowledge work, vision and scientific research. Anthropic notes that the model sustains longer autonomous runs, and its advantage widens as tasks become more complex and extended. Mythos 5, by contrast, is offered through the Project Glasswing trusted‑access program, which currently includes IBM and a handful of critical‑infrastructure operators. The company claims Mythos 5 possesses the strongest cybersecurity capabilities of any AI model to date and will be rolled out gradually to additional partners.

The dual‑release strategy reflects a growing industry trend: separating raw capability from access controls. Gabe Goodhart, chief architect of AI Open Innovation at IBM, told IBM Think that “the launch was inevitable sooner rather than later, since other labs are starting to catch up in capability.” Anthropic’s move positions it as the first to hit this intelligence threshold with a publicly reachable model while keeping the most powerful variant under tighter supervision.

Safety mechanisms differ between the two tiers. Fable 5 incorporates prompt‑level filters, usage‑policy enforcement, and a “dangerous‑task” detector that blocks queries related to weaponization, illicit chemistry or advanced hacking techniques. Mythos 5 adds a second layer: real‑time monitoring by Anthropic’s internal security team and a restricted API that limits batch‑size and request‑frequency, reducing the risk of large‑scale exploitation. Access to Mythos 5 is granted only after partners sign a stringent data‑use agreement and undergo a security audit.

Anthropic’s announcement arrives amid heightened regulatory scrutiny. The U.S. government has recently placed bans on certain high‑risk AI releases, yet the company proceeded, arguing that the added guardrails satisfy emerging policy expectations. Industry observers note that the model’s performance gains—especially in autonomous code generation and scientific literature synthesis—could accelerate adoption in enterprise settings, provided the safety envelope holds.

The release also coincides with notable talent shifts. Nobel laureate John Jumper, co‑creator of AlphaFold, left Google DeepMind to join Anthropic, underscoring the startup’s growing appeal to top researchers. Jumper’s move may signal Anthropic’s intent to deepen its focus on AI‑driven scientific tools, an area where Fable 5 already shows strong benchmark results.

From a practitioner’s perspective, the key question is whether the new guardrails impede legitimate use cases. Early adopters report that the safety filters occasionally flag benign prompts—e.g., detailed queries about enzyme kinetics—requiring workarounds such as rephrasing or invoking the trusted‑access channel. For organizations that need unrestricted access to cutting‑edge reasoning, Mythos 5 remains the only viable path, albeit with higher onboarding overhead.

Historically, AI labs have toggled between open releases and closed‑beta programs. Anthropic’s bifurcated approach may become a template: a “public‑first” model that showcases state‑of‑the‑art performance, paired with a “trusted‑partner” tier that carries the most potent capabilities under strict oversight. If successful, this could reconcile the tension between rapid innovation and societal risk management.

Looking ahead, the industry will watch how Anthropic’s safety layers perform under real‑world pressure. Will Mythos 5’s monitoring catch sophisticated attempts to weaponize the model, or will adversaries find ways to bypass the controls? The answer will shape the next wave of policy and technical safeguards.

FAQ

What is the performance gap between Fable 5 and previous Claude models? Fable 5 leads on nearly all public benchmarks, with especially large margins on long‑running tasks such as multi‑step code synthesis and scientific literature review.

How does Project Glasswing grant access to Mythos 5? Partners must apply, undergo a security audit, sign a data‑use agreement and receive API credentials that enforce rate limits and usage monitoring.

Are the safety guardrails configurable by users? The filters are baked into the model service; users can only request exemptions through the trusted‑access program, which reviews each case manually.

Can I run Fable 5 locally? No. Both models are offered as cloud‑hosted APIs; Anthropic has not released weights for on‑premise deployment.