Claude Fable 5.1 and Mythos 5.1: What Anthropic's New Models Actually Change

Anthropic has announced Claude Fable 5.1, a new flagship model aimed at coding, research, knowledge work, and long-running agentic tasks, alongside Claude Mythos 5.1, a restricted-access version built for vetted professionals working in areas such as cybersecurity and life sciences. As someone who builds a fair amount of my own workflow around Claude Code these days, this is one of those releases where the headline number matters less to me than the quieter change underneath it, how much cheaper it now is to keep an agent working on the same problem for a long time.

The two models share the same underlying capabilities but differ in how their safeguards and access controls are applied. Fable 5.1 is broadly available through Anthropic's products and cloud partners, while Mythos 5.1 is being offered through trusted-access programmes.

What is Claude Fable 5.1?

Claude Fable 5.1 is positioned as Anthropic's latest general-purpose model for demanding technical and professional workflows. The company says it improves on Claude Fable 5 in software engineering, computer use, multi-step reasoning, knowledge work, scientific research, and long-running agentic tasks. Anthropic also says the model is designed to focus more consistently on the underlying cause of technical problems rather than taking shortcuts, which is exactly the kind of behaviour that matters most once you start trusting a model with longer, less supervised chains of work.

Developers can access Fable 5.1 through Claude's platform as well as Amazon Web Services, Google Cloud, and Microsoft Azure. The API model identifier is claude-fable-5-1.

Lower costs for long-running workloads

One of the biggest changes with Fable 5.1 is the reduction in cache-read pricing. Cache reads let a model reuse context it has already processed instead of reprocessing the same information from scratch, and Anthropic says cache-read pricing has been cut by 75 percent, down to $0.25 per million tokens. Input tokens remain priced at $10 per million, and output tokens remain at $50 per million tokens.

Usage Fable 5.1 pricing / claimed reduction

Input $10 per million tokens

Output $50 per million tokens

Cache reads $0.25 per million tokens

Typical workloads About 25% lower

Highly agentic workloads Up to about 45% lower

According to Anthropic, the lower cache-read cost could reduce typical workloads by around 25 percent, while highly agentic coding and research workloads could see savings of up to approximately 45 percent. That matters a lot for anyone building agents that repeatedly work with the same codebase, documents, or research context, which describes most of what I actually use Claude for day to day.

Anthropic's reported benchmark results

Anthropic says Fable 5.1 delivers improvements across coding, computer-use, research, and business workflows. The company's published results include 55.8 percent on Terminal-Bench 4.0, 73.4 percent on CursorBench 3.2.0, 77.9 percent on OSWorld 2.0's partial computer-use measure, 60.9 percent on Humanity's Last Exam without tools and 65.0 percent with tools, 31.4 percent on AutomationBench, and 52.6 percent on Terminal-Bench-Science 0.1.

These are Anthropic-reported results rather than independent evaluations, so they're worth reading within the company's own testing methodology rather than treating as a neutral third-party benchmark.

Claude Mythos 5.1: more restricted access

Alongside Fable 5.1, Anthropic has introduced Claude Mythos 5.1 for professionals working on higher-risk technical and scientific tasks. The underlying model capabilities are shared with Fable 5.1, but Mythos operates under different access controls and safeguards, offered initially through a Cyber Verification Program for approved cybersecurity professionals doing defensive security work, and a Life Sciences Verification Program for approved researchers conducting professional life-sciences research and development.

Anthropic says the life-sciences programme was developed in partnership with the US government, with initial access limited to a group of US organisations and broader access planned later. Anthropic's Claude Security product, which scans codebases for vulnerabilities and proposes patches for human review, is also being powered by Mythos 5.1.

Updated safeguards for cybersecurity

Anthropic says it has changed its safeguards to reduce unnecessary interruptions when users are working on legitimate cybersecurity and biology-related tasks. For cybersecurity, the company says Fable 5.1 can help identify software vulnerabilities while continuing to restrict certain higher-risk activities involving exploit development, penetration testing, and binary-based vulnerability scanning.

Anthropic estimates that Claude Code users will experience approximately 60 percent fewer cybersecurity safeguard interventions per session compared with the previous Fable 5 safeguards. The goal is to make legitimate security work less frustrating without removing restrictions around higher-risk capabilities, and anyone who has hit an unnecessary refusal mid-task will understand why that number matters.

New enterprise privacy controls

Anthropic is also introducing Enterprise Frontier Safeguards, or EFS, a system designed to let enterprises keep their data within infrastructure they control while still providing mechanisms for detecting and responding to potential misuse. Anthropic says EFS will begin rolling out in phases later in the autumn, and in the meantime, eligible enterprise customers can access Fable 5.1 with zero data retention. For organisations handling sensitive business information, these controls could become an important part of how AI agents get deployed inside corporate environments.

AI-assisted scientific research

Anthropic has highlighted several examples of Fable and Mythos being used for scientific research. One demonstration involved Claude Mythos 5.1 designing high-affinity protein binders using open-source protein-design and folding tools, with Anthropic reporting a nearly 50 percent hit rate across 12 targets, compared with what it described as typical protein-design hit rates of around 10 to 15 percent.

In another demonstration, Fable 5.1 was used to train a neural network designed to create a higher-resolution elevation map of approximately one-third of Venus using radar imagery from NASA's Magellan mission. Anthropic said the approach improved detail from approximately 10 to 20 kilometres down to 2 to 3 kilometres, and improved height accuracy by up to 25 percent. These are notable demonstrations, but the results remain company-reported claims and should be independently evaluated by relevant scientific and technical communities before anyone treats them as settled.

Invisible AI watermarking

Anthropic is also introducing an invisible text watermark for outputs from models released after August 2, 2026, following its commitments under the EU AI Act's Code of Practice on Transparency of AI-Generated Content. The watermark is designed to remain invisible in ordinary text and, according to Anthropic, does not contain information identifying individual users, organisations, or conversations. The company is also developing a detection API in private preview for eligible organisations, including regulators, researchers, media organisations, fact-checkers, educational institutions, and certain enterprises with compliance requirements.

What this means for developers

For developers, the biggest change here isn't simply a higher benchmark score. The combination of stronger agentic performance and cheaper context reuse could make Fable 5.1 particularly attractive for long-running AI workflows. A coding agent, for example, may repeatedly reference the same repository, documentation, and previous results during a lengthy task, and lower cache-read costs make those workflows meaningfully cheaper to run. The same principle applies to research assistants, large-document analysis, and business agents that need to hold substantial context over extended sessions.

What it means for enterprises and researchers

For businesses, this announcement puts a lot of emphasis on privacy, data retention, and controlled AI deployment. For cybersecurity and life-sciences professionals, Mythos 5.1 represents a different approach, providing access to more capable systems for specialised work while restricting access to vetted organisations and professionals. That distinction reflects a broader challenge every AI company is navigating right now, making increasingly capable models useful for legitimate high-value work while keeping safeguards around potentially dangerous applications intact.

Where this is heading

The launch of Claude Fable 5.1 and Mythos 5.1 points to where Anthropic appears to be taking its models next, longer-running, more capable AI systems that can carry out complex tasks with less constant human intervention. For everyday users, that shows up as better coding, research, and knowledge-work capabilities. For developers, cheaper context reuse makes autonomous workflows more economical to run. And for enterprises and specialised researchers, Anthropic is placing more emphasis on controlled access, privacy, and tailored safeguards.

What I take from this release, working with these tools daily, is that the strategy has moved past making a chatbot smarter. It's about making a system capable of sitting with you on a complex task for hours instead of minutes, and increasingly, working alongside people for much longer stretches of time.