On September 1, 2026, Anthropic officially released the next-generation flagship models Claude Fable 5.1 and Claude Mythos 5.1. Both share the same base weights but have different safety guard settings: Fable 5.1 is aimed at broad users, while Mythos 5.1 is available only to vetted cybersecurity and life-sciences research institutions through the trusted-access program (Project Glasswing). The company says the new models set a new standard for coding, knowledge work and long-horizon problem solving, with significant metric gains across benchmarks, while cache price reductions and guard refinements lower real-world costs and false-positive rates.

Details: dual-version design and core upgrades

Fable 5.1 and Mythos 5.1 are Anthropic's "Mythos-class" capability models, positioned to handle the most complex, long-horizon knowledge work and coding tasks. Key specs include:

Context window: 1,000,000 tokens
Maximum output: 128,000 tokens
Pricing: $10 input per million tokens, $50 output per million tokens (unchanged from Fable 5)
Cache read price sharply reduced to one-quarter of the previous level, making complex agent task costs roughly 25%–45% lower

Performance improvements: Official and third-party evaluations show Fable 5.1 outperforms its predecessor Fable 5 in coding, long-horizon agent tasks and knowledge work, setting records on benchmarks such as Terminal-Bench and Humanity's Last Exam. Platforms like Vals.ai report a leading Vals Index score, with strong results on LiveCodeBench, ProofBench and others. The model better avoids superficial shortcuts, addresses problems at their source, and maintains readability and consistency across multi-step tasks.

Safety guard refinements: Fable 5.1's cybersecurity and biology-related guards are more precise, with false positives significantly reduced—the company says interventions on benign biology requests are about 85% lower than with Fable 5. The model can be used more effectively for source-code vulnerability identification while still restricting high-risk penetration testing and exploit generation. Mythos 5.1 offers more permissive access to vetted defenders and life-science researchers.

Privacy and deployment: Anthropic is advancing "zero data retention" and Enterprise Frontier Safeguards, allowing enterprise customers to run the models on their own cloud infrastructure so data does not leave premises. This high-privacy service will roll out progressively in the fall.

The models are available via the Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and other platforms, supporting Pro, Max, Team and Enterprise customers.

Technical and market background

Since the Fable series launched in June 2026, it has been positioned as Anthropic's strongest publicly available model, emphasizing long-horizon agent capabilities and complex task handling. Version 5.1 keeps pricing unchanged while responding to user feedback on cost and over-restriction through cache optimization and guard refinement. The Mythos variant continues the "high-capability + strict access" strategy for sensitive research.

Anthropic also showcased pre-release scientific discovery examples generated by the model (such as custom GPU optimizations and high-resolution Venus map reconstruction), attempting to demonstrate preliminary research assistance value.

Potential impact: costs, availability and competitive landscape

1. Impact on enterprises and developers

The cache price reduction lowers costs for long-horizon agent workflows (multi-round tool calls, persistent context), aiding large-scale deployment of coding assistants, research agents and office automation scenarios. More precise guards r