Anthropic has officially announced Claude Fable 5 and Claude Mythos 5, the latest additions to its AI model portfolio. The new models introduce a higher-performance Mythos-class tier above the existing Opus family, delivering substantial improvements in software engineering, scientific research, visual understanding, long-context processing, and advanced reasoning.

Claude Fable 5 is positioned as Anthropic’s most capable generally available AI model to date, while Claude Mythos 5 is a restricted-access version designed for approved cybersecurity defenders, infrastructure partners, and scientific research organizations operating in controlled environments.

The announcement follows the release of Claude Mythos Preview earlier this year under Anthropic’s Project Glasswing initiative.

One of the biggest highlights of Claude Fable 5 is its ability to handle increasingly complex tasks across coding, research, and knowledge-intensive workflows. Anthropic says performance gains become more pronounced as task complexity increases, particularly on long-running, multi-step projects.

Major Improvements in Software Engineering

Anthropic says Claude Fable 5 delivers significant gains in software engineering workloads, including large-scale migrations, codebase refactoring, debugging, architecture planning, and production system modifications.

According to early testing, the model can perform engineering tasks that previously required months of effort in a significantly shorter timeframe.

The company highlighted an internal evaluation involving Stripe, where Claude Fable 5 reportedly compressed months of development work into days, including migration tasks spanning a 50-million-line Ruby codebase.

Fable 5 also demonstrates stronger performance across production-grade coding benchmarks and software engineering evaluations, including Cognition’s FrontierCode benchmark.

Enhanced Reasoning & Knowledge Work

Beyond coding, Claude Fable 5 introduces improvements in structured reasoning, financial analysis, document understanding, chart interpretation, and multi-step decision-making.

Anthropic says the model achieved leading performance on Hebbia’s Finance Benchmark, demonstrating strong capabilities in tasks typically handled by senior analysts and knowledge workers.

Additional evaluations conducted by IMC reported improvements in:

• Financial modeling
• Trading analysis
• Root-cause investigations
• Expected-value calculations
• Conceptual reasoning
• Fact-based information retrieval

Improved vision and visual reasoning

Claude Fable 5 also expands Anthropic’s multimodal capabilities.

The model can extract detailed information from:

• Scientific charts
• Research papers
• Technical diagrams
• Complex tables
• Screenshots
• Software interfaces

Anthropic says the model requires significantly less external tooling than previous generations and can accurately reconstruct application code directly from screenshots.

In experimental testing, Claude Fable 5 successfully completed gameplay tasks in Pokémon FireRed using only visual input, demonstrating stronger visual understanding and planning abilities than earlier Claude models.

Long-Context Memory & Persistent Reasoning

Memory and context management represent another major area of improvement.

Claude Fable 5 supports multi-million-token context windows and introduces enhanced persistent memory capabilities that allow the model to maintain coherence throughout lengthy workflows.

The system can generate and reference self-created notes during long-running tasks, improving consistency and task completion quality.

Anthropic reports that persistent memory improved performance significantly during complex gaming evaluations such as Slay the Spire, where the model demonstrated much stronger long-term planning capabilities compared to previous generations.

Claude Mythos 5 Targets Scientific Research

While Claude Fable 5 is designed for broad availability, Claude Mythos 5 focuses on advanced scientific and cybersecurity research.

The model operates with reduced safeguards under tightly controlled access programs and is currently available only to trusted organizations and approved research partners.

According to Anthropic, Mythos 5 demonstrates advanced capabilities across:

• Protein engineering
• Drug discovery workflows
• Bioinformatics
• Computational biology
• Genomics research
• Autonomous scientific investigation
• Multi-step laboratory planning

The company says Mythos 5 successfully completed end-to-end scientific workflows with minimal human supervision and generated several novel biological hypotheses that have advanced to experimental validation.

Among its reported achievements are:

• Discovery of new molecular biology hypotheses
• Research involving E. coli protein mechanisms
• Single-cell genomics studies spanning 138 animal species
• Machine learning models identifying equivalent cell types across distant species
• Research outputs that reportedly outperform some recently published scientific models

Several findings are currently undergoing further evaluation, while others are expected to be published in future research papers.

Safety Systems & Alignment

Anthropic says both models incorporate expanded safety systems designed to balance capability with responsible deployment.

Claude Fable 5 includes AI-based safety classifiers that automatically detect potentially sensitive requests involving:

• Cybersecurity
• Biology
• Chemistry
• Model distillation attempts

When necessary, requests may be routed to safer Claude models instead of Fable 5.

The company has also implemented enhanced protections against offensive cybersecurity activities, including:

• Vulnerability exploitation
• Multi-stage attack planning
• Reconnaissance workflows
• Lateral movement guidance

Additional safeguards target sensitive biological and chemical research use cases due to the increased scientific capabilities demonstrated by the Mythos-class architecture.

Anthropic says internal testing found low levels of deceptive behavior and misuse cooperation, with alignment characteristics comparable to previous Opus-class models.

Data Retention & Privacy

For Mythos-class deployments, Anthropic applies a 30-day retention policy.

The company says:

• Data is not used for model training
• Access is logged and monitored
• Data is typically deleted after 30 days
• Retention exists solely for safety monitoring and abuse prevention

Claude Fable 5 & Claude Mythos 5 Price & Availability

Both models are priced at:

• $10 per million input tokens
• $50 per million output tokens

Claude Fable 5

• Available through Anthropic API
• Available to Pro subscribers
• Available to Max subscribers
• Available to Team plans
• Available to Enterprise customers
• Included at no additional cost during the initial rollout period
• Will transition to a usage-credit model later based on infrastructure capacity

Claude Mythos 5

• Restricted access
• Available only to approved cybersecurity organizations
• Available to selected scientific research partners
• Expansion planned through controlled access programs for cybersecurity and biomedical research

Anthropic says Claude Fable 5 represents its most capable generally available AI system so far, while Claude Mythos 5 serves as a preview of how increasingly powerful AI systems could accelerate scientific discovery and cybersecurity research under carefully controlled conditions.