Anthropic Launches Claude Fable 5.1 and Mythos 5.1: Frontier AI for Coding, Cyber Defense and Science

Posted on 3rd Sep 2026 06:04:21 in Artificial Intelligence, Machine Learning

Tagged as: Anthropic, Claude, AI Models, Cybersecurity, Generative AI

Anthropic has shipped its most ambitious release of 2026. On September 1, the company introduced Claude Fable 5.1 and Claude Mythos 5.1, describing them as "the world's most advanced models for coding and knowledge work." The twist: both names point to the same underlying model, offered at two different safeguard levels. Fable 5.1 is generally available to every customer, while Mythos 5.1 is reserved for vetted cyber defenders and life scientists through Anthropic's trusted access programs. The launch lands the same week that Google and OpenAI have also expanded their cyber-focused AI programs, signalling how fast frontier labs are moving to make AI a working tool for security teams and researchers.

One Model, Two Safeguard Levels

The Fable and Mythos split is Anthropic's answer to a hard problem: how do you give security and biology researchers the full power of a frontier model without handing attackers the same capability? Both versions share identical weights, but Mythos 5.1 keeps its complete cyber and bio capabilities intact for defensive work, vulnerability research and drug design. Fable 5.1, the broadly available edition, is tuned for enterprise coding, long-running agentic projects and knowledge work.

The raw specifications are unchanged from the Fable 5 line: a 1 million token context window, 128K maximum output tokens, adaptive thinking that is always on, and a knowledge cutoff of June 2026. Developers can call the model today as claude-fable-5-1 on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Azure. The company says Fable 5.1 defaults to High effort in Claude Code and Medium effort in Claude Cowork and on Claude.ai, which means most users get a meaningful capability bump without changing a single setting.

Benchmarks: A Clear Step Up Across the Board

Anthropic's numbers show Fable 5.1 delivering its biggest generational jump yet. On Terminal-Bench 4.0, the agentic coding benchmark, Fable 5.1 scores 55.8 percent against 42.0 percent for Fable 5, 52.3 percent for Opus 5 and 37.3 percent for GPT-5.6 Sol; Mythos 5.1 reaches 60.9 percent. On Humanity's Last Exam, the flagship multidisciplinary reasoning test, it hits 65.0 percent with tools compared with 63.8 percent for Fable 5. On CursorBench 3.2.0 it posts 73.4 percent, the highest score recorded on the benchmark at launch, and on OSWorld 2.0 it completes 77.9 percent of computer-use tasks under partial credit scoring.

The most striking gain is in scientific agentic work: on Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6 percent, more than double Fable 5's 24.7 percent. That leap is why early-access partners describe the model as qualitatively different for long-horizon work. At Millennium, testers reported that Fable 5.1 found the cause of a rare, one-in-a-million crash in internal systems that no engineer or prior model had explained in years, disassembling an external vendor library and matching it against a core dump. Cognition said it is moving its Opus 5 traffic in Devin to Fable 5.1 on launch day. MongoDB engineers reported a three-day unattended run that produced a working prototype with verification loops, and Ramp described a 38-hour unattended machine learning run that diagnosed a label artifact, corrected it and returned results with next steps.

Cheaper to Run, Safer to Operate

Cost was a major source of customer feedback on Fable 5, and Anthropic responded directly. Cache reads now cost USD 0.25 per million tokens, a 75 percent reduction, which cuts the cost of typical workloads by roughly 25 percent and highly agentic workloads by up to 45 percent. Input and output pricing otherwise stays at USD 10 and USD 50 per million tokens.

On the trust side, the company introduced Enterprise Frontier Safeguards (EFS), a system that stores customer data in cloud infrastructure controlled entirely by the customer rather than Anthropic. EFS rolls out to enterprise customers in phases beginning later this fall; until then, eligible customers can run Fable 5.1 with zero data retention. Anthropic has also designated Fable 5.1 a Covered Model, a category that carries additional data retention, safety review and access policies wherever the model is offered. Cyber safeguards now produce 60 percent fewer false positives, and Fable 5.1 can be used to discover software vulnerabilities, though explicitly not to develop exploits for them. The release is also EU AI Act compliant, with the invisible text watermark that became mandatory for models released after August 2, 2026, and a detection API now in private preview for regulators, media and fact-checkers.

From Protein Binders to a Map of Venus

The science results in the launch post hint at where this model family is heading. Given access to open-source protein design tools, Mythos 5.1 designed high-affinity protein binders that external labs confirmed experimentally: on three competition targets its designs showed roughly 10 times higher binding affinity than the best human submissions, and across 12 targets it achieved a hit rate near 50 percent, far above the 10 to 15 percent typical in protein design today.

In a separate demonstration, Fable 5.1 trained a neural network to build a new elevation map of a third of Venus from Magellan mission radar images captured more than 30 years ago. The map resolves details down to two to three kilometres, compared with 10 to 20 kilometres previously, and measures heights up to 25 percent more accurately. Anthropic has released the map under a Creative Commons licence ahead of NASA's VERITAS and ESA's EnVision missions, so mission planners can use it to pick geological targets.

What This Means for Developers and Security Teams

For software teams, Fable 5.1 is the strongest argument yet that long-running agentic coding has moved from demo to production: sessions that run for hours, own entire features across codebases and report honestly when stuck rather than papering over failures. For security teams, the Mythos tier plus the cyber safeguard improvements mean defensive tooling is being treated as a first-class product rather than an afterthought. And for enterprises watching cost curves, the cache-read repricing is the clearest signal that frontier-model economics are improving even as capability climbs.

The competitive picture is equally important. With Fable 5.1 topping GPT-5.6 Sol on every agentic benchmark Anthropic published, and Google expanding its own cyber model lineup in the same week, the frontier race is now running on three tracks at once: raw intelligence, agentic reliability and trusted access. Claude Fable 5.1 is available now; cyber defenders and life scientists can register interest in Mythos 5.1 through Anthropic's access programs.

Sources

• Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1
• AWS — Claude Fable 5.1, Anthropic's New Frontier Model Is Now Available on AWS
• The Hacker News — Google, Anthropic, and OpenAI Unveil Cyber AI Models, Safeguards, and Access Programs
• Claude Platform Docs — Claude Fable 5.1 Overview and Specifications