Claude Fable 5 and Mythos 5: SOTA Autonomy and Safeguards

TL;DR: Anthropic officially launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Claude Fable 5 is a general-use, Mythos-class model featuring SOTA agentic autonomy in coding, vision, and biology. Claude Mythos 5 is the same model but has safeguards lifted for trusted defense partners under Project Glasswing. Both models are available today at $10 per million input tokens and $50 per million output tokens, representing a pricing drop of more than 50% compared to preview versions.

Claude Fable 5 & Mythos 5: SOTA Autonomy and Safeguards

On June 9, 2026, Anthropic announced the launch of Claude Fable 5 and Claude Mythos 5, marking the arrival of the Mythos-class models for general use. These models sit above the Opus class in overall reasoning and represent a major leap forward in autonomous agentic capabilities.

The release resolves the prediction market speculation we analyzed in Anthropic Mythos Public Release: Polymarket Odds Surge to 92%—confirming the public availability of Anthropic's new flagships.

Fable 5 and Mythos 5 are available immediately, offering a significant pricing drop compared to the preview models:

• Input: $10 per million tokens
• Output: $50 per million tokens

Note: This pricing is less than half the cost of the previous Claude Mythos Preview model ($25/$125).

The Etymology: Fable vs. Mythos

The distinction between the two models lies in their safety configurations. Fable is derived from the Latin fabula, meaning "that which is told," akin to the Greek mythos.

While Claude Mythos 5 represents the raw, unconstrained model deployed in high-security settings, Claude Fable 5 is the same underlying model wrapped in a cautious safety classifier system designed for general public use.

Technical Performance & Partner Feedback

The core strength of the Mythos-class models is their ability to work autonomously for longer periods than any previous Claude models. Anthropic reports that the longer and more complex the task, the larger Fable 5’s lead over other models.

1. Software Engineering & Agentic Coding

Fable 5 exhibits exceptional performance in software engineering, and is more token-efficient than prior models.

• Stripe reported that Fable 5 compressed months of engineering into days, performing a codebase-wide migration on a 50-million-line Ruby codebase in a single day—a task that would have taken a team of developers over two months by hand.
• Luke Anderson (CTO) noted that Fable 5 handles the complex multi-agent workflows that developers run daily in Claude Code, delivering more capable engineering in fewer turns.
• On Cognition's FrontierCode evaluation (FrontierBench), which measures model performance on production-grade codebases, Fable 5 scored the highest among frontier models. CEO Scott Wu shared that the model generalizes to unfamiliar tools out of the box and excels at long-horizon reasoning.

2. High-Level Knowledge Work & Finance

On complex analytical tasks, Fable 5 shows substantial gains:

• Hebbia's Finance Benchmark for senior-level reasoning: Fable 5 has the highest score of any model, excelling at document-based reasoning, chart and table interpretation, and problem solving.
• IMC trading analysis: Fable 5 aced trading evaluations nearly across the board, including factual lookup, expected-value, conceptual reasoning, and root-cause analysis.
• Izzy Miller (AI Research Lead) stated that Fable 5 is the first model to break 90% on their core analytics benchmark of complex, long-running analytical tasks, showing a 10-point jump over Opus.

3. Frontier Physics and Scientific Research

• Matthew Pines (CEO) shared that Fable 5 is the strongest model they have tested on frontier physics research. Using only a third of the reasoning tokens, Fable 5 reached a level of progress in 36 hours that took competing models like GPT-5.5 four days to achieve.

4. Vision-Only Autonomy

• Fable 5 sets a new state-of-the-art for visual processing and multi-modal tasks. It can extract precise numbers from detailed scientific figures and rebuild a web application's source code directly from screenshots alone.
• Previous Claude models struggled to play Pokémon FireRed even with complex helper harnesses that gave them additional maps and navigation aids. Fable 5 completed the game using a minimal, vision-only harness, operating purely off raw game screenshots.

5. Memory & Long-Context Horizon

• Fable 5 stays focused across millions of tokens in long-running tasks. When playing the deck-building game Slay the Spire, giving the model access to persistent file-based memory improved its performance three times more than it did for Opus 4.8, allowing it to reach the game's final act three times more often.

6. 3D Worldbuilding and Game Development

Fable 5 has demonstrated remarkable capabilities in autonomous 3D world creation and game development:

• Minecraft Clone Generation: In real-world tests, Fable 5 created fully functional Minecraft clones in a single 20-55 minute session, complete with multiple biomes, day/night cycles, cave systems, different ore types, and procedural terrain generation—all built from scratch using custom ThreeJS code.
• Browser-Based 3D Worlds: Developers like Matt Shumer showcased Fable 5 building complex 3D voxel worlds running entirely in the browser using custom-built ThreeJS implementations, with users able to optimize frame rates on command with simple prompts like "make it faster, without losing quality."
• NASA Data Integration: In one demonstration, Fable 5 created a to-scale Yosemite Valley environment with over 266,000 trees and realistic waterfalls using real NASA topographical data, showcasing its ability to integrate external datasets into 3D visualizations.
• Iterative Performance Optimization: When users reported performance issues, Fable 5 autonomously refactored rendering pipelines to improve frame rates while maintaining visual quality, demonstrating real-time problem-solving in graphics programming.

This capability represents a significant advancement in "vibe-coding" for 3D prototyping, allowing developers to create complex interactive environments without heavy game engines or extensive graphics programming expertise. For more details on Fable 5's game development capabilities, see Claude Fable 5 Creates Minecraft Clones and 3D Worlds from Simple Prompts.

Scientific Breakthroughs: Biology & Genomics

Using Claude Mythos 5, researchers have unlocked significant accelerations in life sciences research.

Autonomous Drug Design

Anthropic's internal protein design experts used Mythos 5 to accelerate aspects of the drug design process by approximately 10 times. Operating autonomously with standard protein design and bioinformatics tools (and without human assistance), Mythos 5 matched or outperformed skilled human scientists. The model successfully identified binding sites, selected molecular tools, and recovered from errors. Nine of the 14 protein targets from this study (such as immune checkpoints and growth-factor receptors) yielded strong candidates that are currently under experimental evaluation.

Gene Therapy and Virus Assembly

• Evaluated on predicting the properties of the viral shell of adeno-associated viruses (AAVs) developed by Dyno Therapeutics, Mythos-class models outperformed dedicated protein language models using biological reasoning alone, despite having no explicit training on the task. AAVs are critical components for delivering gene therapies, making this a promising capability for bio-research.

Genomics and Model Training

• In a week of largely autonomous work, Mythos 5 assembled single-cell genomic data for millions of cells across 138 animal species. It designed and trained a custom machine learning model to identify cells performing identical roles in distantly related organisms. Despite being 100 times smaller, Mythos 5's trained model outperformed a recent model published in the journal Science.

Deep Dive: Fable 5’s Safeguard Framework

Because a model with Mythos-level capabilities poses a risk of uplift to malicious actors (such as providing dangerous advice on cybersecurity or biochemistry), Anthropic has wrapped Fable 5 in a new safety framework.

graph TD
    User([User Prompt]) --> Filter{Safeguard Triggered?}
    Filter -- "Yes (under 5% of sessions)" --> Fallback[Claude Opus 4.8 Response]
    Filter -- No --> Fable[Claude Fable 5 Response]

Safety Classifiers and the Opus 4.8 Fallback

• Safety Classifiers: Fable 5 uses separate AI systems to detect potential misuse and prevent the main model from responding.
• The Opus Fallback: When classifiers flag a request related to cybersecurity, biology/chemistry, or distillation (attempts to extract model capabilities), the response is automatically handled by Claude Opus 4.8 instead of issuing an outright refusal.
• Trigger Rate: Because safeguards are tuned conservatively, they may catch harmless requests; however, they trigger in less than 5% of user sessions on average. More than 95% of Fable sessions involve no fallback.

Red-Teaming and Jailbreak Resistance

• Jailbreak Testing: Anthropic ran an external bug bounty program that produced no universal jailbreaks in over 1,000 hours of testing. External red-teaming organizations and the UK AISI also tested the model on long-form agentic tasks.
• Classifier Robustness: In external testing, Fable 5 complied with zero harmful requests relating to cyberattacks, exploit development, or defense evasion, even when subjected to 30 different public jailbreak techniques.

Data Retention Policy

To defend against complex, multi-request attacks and jailbreaks, Anthropic is introducing a new data retention policy for Mythos-class models:

• 30-Day Retention: All traffic on Mythos-class models (first- and third-party surfaces) will be retained for 30 days.
• No Training: This data will not be used to train new Claude models or for any non-safety purpose.
• Privacy Protections: All human access to this logs database is audited, and data is deleted after 30 days in almost all cases.

Subscription Schedule & API Availability

Due to high demand and capacity constraints, Anthropic is rolling out access to subscription plans in stages:

• June 9 – June 22: Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
• June 23: Fable 5 will be removed from subscription plans. Continued use will require usage credits.
• Future Restore: Once capacity allows, Anthropic intends to restore Fable 5 as a standard part of subscription plans.
• API Availability: Developers can access Fable 5 today via the endpoint claude-fable-5.

Example API Request (Python)

import anthropic

client = anthropic.Anthropic(api_key="your-api-key")

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=2048,
    messages=[
        {"role": "user", "content": "Analyze this codebase screenshot and reconstruct the React layout."}
    ]
)

print(response.content[0].text)

Sources & References

• Anthropic News: Claude Fable 5 and Claude Mythos 5 Launch Announcement
• Safety & Alignment: Claude Fable 5 & Mythos 5 System Card
• Project Glasswing: Initial Project Glasswing Update

Related Reading

• Claude Fable 5 Creates Minecraft Clones and 3D Worlds from Simple Prompts
• Self-Harness: AI Agents That Improve Their Own Framework
• Claude Opus 4.8: Agentic Improvements & Benchmarks
• Claude Mythos Preview and Cybersecurity Analysis
• What is Claude.md? Persistent Memory for AI Agents

Claude Fable 5 and Mythos 5 were launched by Anthropic on June 9, 2026. Fable 5 is generally available for developers via API and web interface, while Mythos 5 is gated for cyberdefenders via Project Glasswing. Pricing is set at $10/M input and $50/M output tokens.