On June 9, Anthropic released the public version of Claude Fable 5 and the restricted version of Claude Mythos 5. As the first Mythos-class flagship models, they set new SOTA benchmarks in software engineering, multimodal vision, and life sciences. Both models use an AI classifier to automatically downgrade the safety guardrails, involving high-risk automatic switching to Claude Opus 4.8.
Claude Fable 5 Performance Data: Three Officially Confirmed Benchmark Results
The evaluation data published by Anthropic is as follows：
Software engineering：In a real test on Stripe involving a migration project with 50 million lines of Ruby code, the engineering team originally needed two months; Claude Fable 5 completed the refactor within one day
Multimodal vision：Can restore and reconstruct code from web screenshots with 100% fidelity; can clear Pokémon FireRed autonomously using only the original images as input, with no assistive tools
Life sciences (Mythos 5限定)：Drug design speed increased by 10x; the blind test preference rate for scientists reached 80%
In an official statement, Dianne Penn, Anthropic’s head of product management, confirmed that the Token consumption required to execute tasks with Fable 5 is lower than that of the prior generation model, and the overall cost per task actually decreases.
Automatic Downgrade Guardrail Confirmation Specs: Trigger Conditions and Impact Scope
According to Anthropic’s official data, the technical specs for confirming the automatic downgrade guardrails are as follows：
Trigger conditions (any of the following)：
· Highly sensitive network security attack requests
· Synthesis of biological or chemical hazard substances
· Malicious model distillation attempts
Handling method： The system switches the conversation to Claude Opus 4.8 on the device or in the cloud to provide a response, and proactively notifies the user.
Impact scope： Official data shows that, on average, fewer than 5% of conversation Sessions trigger this mechanism.
Anthropic’s official statement says it will continue to optimize guardrails to reduce false positive rates, without disclosing a specific optimization timeline.
Claude Mythos 5 Access Restrictions: About 200 Institutions, Not Open to the Public
According to Anthropic’s official announcement, the licensed access scope of Claude Mythos 5 is as follows：
Project Glasswing participating institutions：Currently about 200, covering institutions that participate in Anthropic and U.S. government critical infrastructure protection collaborations
Critical infrastructure providers：For specific defense and security defense purposes
Users who hold access permissions for the old Claude Mythos preview (no safety-protection version) can be upgraded to Claude Mythos 5, Anthropic confirmed.
Pricing and Free Trial Schedule
API pricing (same for Fable 5 and Mythos 5)：
Input：$10 per million Tokens
Output：$50 per million Tokens
This pricing is less than half of the Mythos test version pricing.
Free trial period (limited to subscribers)： Pro, Max, Team, and Enterprise subscribers can access Claude Fable 5 for free, limited to June 9, 2026 through June 22.
Access channels： The Claude web version, Claude App, and the developer API are open to users worldwide starting immediately.
Frequently Asked Questions
What are the core differences between Claude Fable 5 and Claude Mythos 5?
Claude Fable 5 is the public-facing version, with built-in automatic downgrade guardrails that automatically switch high-risk queries to Opus 4.8 for handling. Claude Mythos 5 is a version that removes some safety restrictions, open only to roughly 200 authorized institutions under Project Glasswing, and is intended for scenarios such as critical infrastructure defense for government collaborations and regulated drug research.
After the automatic downgrade mechanism is triggered, how does the user’s conversation change?
According to Anthropic’s official description, the system automatically switches to Claude Opus 4.8 to handle the query and proactively notifies the user that a downgrade has occurred. Official data shows the trigger rate is below 5%; ordinary tasks such as general software development, text generation, and data analysis are not within the trigger scope.
After the free trial period for Pro/Max subscribers ends, how are access costs calculated?
After the free trial period ends on June 22, the API pricing for Claude Fable 5 is $10 per million Tokens for input. Anthropic’s official currently has not announced specific change plans for the subsequent access rates for subscribers.
