Claude Fable is Live...and its WILD!

Key Concepts

• Fable 5: The new, generally available, safety-guarded version of Anthropic’s latest "Methus" class model.
• Methus 5: The unrestricted version of the model, currently limited to select organizations under "Project Glass Swing."
• Test-Time Compute (Thinking Budget): The concept of increasing reasoning effort by allocating more tokens to the model's internal processing, leading to higher performance.
• Frontier Code Benchmark: A specialized evaluation suite by Cognition Labs designed to replicate real-world software engineering tasks.
• Safety Classifiers: An architectural layer that intercepts prompts (e.g., cyber-security related) and routes them to a less capable model (Claude Opus 4.8) to prevent misuse.
• Data Retention Policy: A mandatory requirement for Methus-class models to store all traffic for 30 days to improve safety and defense against novel attacks.

---

1. Model Overview and Performance

Anthropic has released its most significant model in three years, categorized as the "Methus" class.

• Benchmark Dominance: In the Cognition Labs "Frontier Code" benchmark, Fable 5 achieved a 29% score, which later rose to 46% on the full extended benchmark. This significantly outperforms Claude Opus 4.8 (13%) and GPT 5.5 (6%).
• Scaling Reasoning: A key finding is that performance scales with "thinking budget." By increasing test-time compute, users can achieve up to 3x the performance of Opus 4.8 at a similar cost, provided the reasoning budget is managed effectively.

2. Safety Architecture and "Project Glass Swing"

Anthropic has implemented a dual-model strategy to balance power with safety:

• Fable 5 (Public): Includes an extra classifier layer. If a user query triggers safety concerns—specifically regarding cyber-security vulnerabilities—the system automatically routes the request to the less capable Claude Opus 4.8.
• Methus 5 (Restricted): The raw, unrestricted model. Access is currently limited to organizations within "Project Glass Swing."
• False Positives: Anthropic acknowledges that the safety classifier may trigger on approximately 5% of sessions, which may lead to perceived performance degradation for some users.

3. Real-World Applications

The model demonstrates capabilities beyond standard text generation:

• Drug Design: In experimental protein design, Methus 5 expedited the research process by approximately 10x.
• Scientific Hypothesis: It is reportedly the first model capable of generating novel hypotheses in molecular biology and genomics.
• Autonomous Loops: Felix (Claude Code team) suggests a shift from "task-based" AI (where a human initiates every step) to "responsibility-based" AI. In this paradigm, the model runs in a continuous loop to monitor and resolve issues (e.g., preventing app crashes) without constant human intervention.

4. Pricing and Data Policy

• Pricing: Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. While expensive, it remains more affordable than the GPT 5.5 Pro tier ($30/$180+).
• Access: Available immediately via API and enterprise plans. Subscription users have access until June 22, 2025, after which usage will require paid credits.
• Data Retention: A controversial new policy requires the retention of all traffic on Methus-class models for 30 days. Anthropic states this is strictly for safety, identifying jailbreaks, and reducing false positives, with human access to this data being logged.

5. Notable Quotes

• On the "Third Era" of AI: "I no longer tell Claude to investigate a particular crash report. It runs in a loop watching every crash report that comes in. Its job is no longer to help me fix a crash. It's to keep our apps from crashing." — Felix, Claude Code Team.
• On Model Capability: "Fable’s capabilities exceed those of any model that we have ever made generally available." — Anthropic Blog.

Synthesis

The release of Fable 5 marks a transition toward "test-time compute" scaling, where the intelligence of a model is no longer static but can be expanded through increased reasoning effort. While the model sets a new state-of-the-art for software engineering and scientific research, its adoption is tempered by high costs, strict data retention requirements, and a safety-first routing mechanism that may limit its utility for certain technical tasks. The shift toward autonomous, loop-based workflows suggests that the next phase of AI development will focus on persistent, agentic oversight rather than simple prompt-response interactions.