Anthropic Claude Fable 5 and Mythos 5 Release

Overview

Anthropic has released Claude Fable 5, a model categorized as a "Mythos class" model. While the full Mythos 5 model is limited to select users in Project Glasswing (primarily high-spending corporate customers with specific use cases), Fable 5 is the version made safe for general availability. Anthropic claims Fable 5's capabilities exceed any previously available general-use model.

Performance and Benchmarks

Claude Fable 5 demonstrates significant improvements over previous models, particularly in specialized technical and professional domains, though gains are uneven across different benchmarks.

Coding and Agentic Tasks

Fable 5 shows a substantial performance increase in agentic coding. On the SWE-bench Pro and Frontier Code (a benchmark from Cognition), Fable 5 more than doubles the score of Opus 4.8, which previously held a score of 13.4.

Legal and General Benchmarks

• Legal Benchmarks: Fable 5 is approximately 30% better than Opus 4.8 and outperforms GPT-5.5.
• General Performance: The model outperforms both Opus 4.8 and GPT-5.5, though it is noted that the gap is less significant in tool-use and computer-use benchmarks compared to the massive leaps in coding.

Pricing and Access Tiers

API Pricing

Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. This is double the cost of Opus 4.8, though it is less than half the price of the earlier Claude Mythos Preview.

Subscription Changes

Access to Fable 5 is included in Pro and Max accounts until June 22. Starting June 23, Fable 5 will be removed from these plans, requiring users to pay via API token pricing. Anthropic intends to restore Fable 5 to subscription plans as quickly as possible, though it is expected to remain exclusive to the Max tier and higher rather than returning to the Pro tier.

Safety Restrictions and Model Switching

Anthropic has implemented aggressive safety classifiers that can trigger a model switch mid-conversation. If a prompt is flagged, the system automatically switches the user from Fable 5 to a different model (such as Opus 4.8) to continue the chat.

Trigger Categories

• Cybersecurity: Blocks content related to malware or attacks.
• Life Sciences/Biology: Extremely sensitive triggers. Even benign queries—such as asking for a breakdown of current Ebola outbreaks and associated risks for the World Cup—can trigger a safety flag and force a model switch.
• Chain of Thought (CoT): The model is designed to block requests attempting to elicit its internal "summarized thinking" or long chain of thought.

Data Retention Policy Shift

Anthropic is introducing a mandatory 30-day data retention policy for all traffic on Mythos class models across both first-party and third-party surfaces. While Anthropic claims this data will not be used to train new models or for non-safety purposes, the policy is intended to allow the company to capture and analyze jailbreak attempts to patch vulnerabilities quickly.