Anthropic has unveiled two new AI models: Claude Fable 5 for public use and Claude Mythos 5 for trusted industry partners. This release aims to enhance their capabilities while addressing security concerns related to their potential misuse in cyberattacks.
The Claude Mythos 5 is being provided to a select group of technology industry partners, many of whom had early access to a previous model due to fears that its features could be exploited for hacking. Meanwhile, Claude Fable 5 will be made available to the general public with significant restrictions in place to prevent its use in cybersecurity-related queries, which will instead redirect queries to an older model, Claude Opus 4.8. This safeguard reflects concerns about potential misuse, including training smaller models from larger ones, known as distillation.
Diane Penn, Anthropic’s head of product management, stated that the company has been refining its approach based on feedback since the initial release. The protective measures will prioritize caution, potentially leading to misdirected queries even when they are innocent. Over time, the goal is to improve the system’s accuracy while maintaining safety.
Alongside the limited release for Project Glasswing partners, select biology researchers will also get access to Claude Mythos 5. Anthropic plans to broaden this accessibility in the future, indicating that they envision an eventual rollout to more users. Their emphasis on controlled access highlights the growing need for enhanced software security as AI continues to advance.
Both new models feature increased performance for tasks requiring software engineering and visual understanding, with a pricing structure set at $10 per million input tokens and $50 per million output tokens, which is higher than Anthropic’s other public AI models. This strategic release reflects ongoing tensions within the tech industry to balance innovation with essential cybersecurity measures.
As both Anthropic and its competitors like OpenAI navigate these challenges, the anticipation grows regarding how effective these safeguards will be in real-world applications. The understanding remains that robust protections are critical to prevent the capabilities of such advanced AI models from falling into the wrong hands.
