Anthropic 今天同时发布了两个模型：Claude Fable 5 和 Claude Mythos 5。 两个模型用的是同一个底座，区别在于 Fable 5 加了一套安全分类器，面向所有用户开...

Baoyu on X: "Anthropic released two models today: Claude Fable 5 and Claude Mythos 5. Both models are based on the same foundation, but Fable 5 includes an additional set of safety classifiers and is available to all users; Mythos 5 removes some safety restrictions and is only available to cybersecurity partners in Project Glasswing. In short, Fable 5 is the 'guarded' version of Mythos. Two months ago, Mythos Preview was locked within about 200 defense agencies, but now ordinary developers can access the same level of capabilities. [Fable 5's Safety Mechanism] Fable 5's safety mechanism is not the traditional 'refuse to answer,' but rather downgrading: when the classifier detects a request involving cybersecurity attacks, biochemical weapons, or model distillation behavior, it automatically switches to Opus 4.8 to respond and informs the user that a downgrade has occurred. According to Anthropic's data, over 95% of conversations do not trigger a downgrade. Anthropic also admits that the classifier is currently set too strictly, which may mistakenly reject normal requests, and they will continue to optimize to reduce false positives. [How Powerful Are the Capabilities] Anthropic lists a bunch of benchmarks, but a few real-world cases can better illustrate the point. Stripe used Fable 5 to perform a full library migration on a Ruby codebase with 50 million lines, completing it in one day, whereas it would have taken an entire team two months. In Cognition's FrontierCode test, Fable 5 achieved the highest score with moderate computational resources, and its token efficiency is significantly better than previous Claude models. In terms of visual capabilities, previous Claude models needed various auxiliary tools to progress in Pokémon FireRed, but Fable 5 completed the game with just the most basic visual interface. It can also directly reconstruct the source code of a web application from a screenshot. In the field of life sciences, Mythos 5 allowed Anthropic's internal protein design experts to accelerate certain parts of the drug design process by about 10 times. In a genomics study, Mythos 5 worked almost entirely autonomously for more than a week, and the model it trained outperformed a model published in Science, while being only one percent of the size. [Pricing and Availability] The API pricing for Fable 5 and Mythos 5 is $10 per million input tokens and $50 per output token. Compared to Mythos Preview's $25/$125, this is a 60% reduction. However, it is twice as expensive as Opus 4.8's $5/$25. Compared to OpenAI's GPT-5.5 ($5/$30), input is twice as expensive, and output is about 67% more expensive. Subscribers should be aware of a time window: from today to June 22, Pro, Max, Team, and enterprise users can use Fable 5 for free. Starting June 23, using Fable 5 will require purchasing additional usage credits. Anthropic says that once production capacity is sufficient, Fable 5 will be restored as a standard feature of the subscription plan, but they have not provided a specific time. API and pay-as-you-go enterprise users are unaffected and can start calling it normally today. [An Often Overlooked Policy Change] Anthropic also announced that starting with Fable 5, all traffic from Mythos-level models will be forcibly retained for 30 days, covering first-party and third-party platforms. Anthropic promises not to use this data to train models, but only for security monitoring, such as detecting new jailbreak attacks and complex attack patterns across requests. However, for enterprise users who value data privacy, this is a change that needs to be evaluated, especially for those who previously chose Anthropic precisely because of its zero-retention policy.

Claude

@claudeai

15h

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.

00:00

5:22 PM · Jun 9, 2026

103.8K

Views

1

0

4

104

3

8

38

2

280

9

119

Read 104 replies