The Arrival of Mythos
Anthropic has finally pulled back the curtain on its Mythos model, a system so potent it was once deemed too risky for public release. Now, it's here—albeit with a suite of new guardrails and a price tag that reflects its advanced capabilities. The release marks a significant milestone in AI development, offering a glimpse into the future of machine intelligence while attempting to mitigate the risks that come with such power.
The Mythos model isn't just another incremental upgrade. It's a leap forward, positioned above Anthropic's Claude Opus in benchmark performance. This isn't just about raw power, though. Mythos is designed to excel in areas that matter most: software engineering, knowledge work, vision, and scientific research. It's a tool that could redefine what's possible in AI-assisted tasks, from coding to complex problem-solving.
But with great power comes great responsibility, and Anthropic seems acutely aware of this. The release includes two distinct versions: Claude Fable 5, available to the public, and Claude Mythos 5, which is being offered privately to Glasswing partners. This dual approach allows Anthropic to balance accessibility with control, ensuring that the most powerful capabilities are in the hands of those who can use them responsibly.
Safety First: Guardrails and Classifiers
One of the most striking aspects of the Mythos release is the emphasis on safety. Anthropic has equipped Claude Fable 5 with new classifiers designed to detect misuse. These classifiers act as a first line of defense, identifying potentially harmful queries and ensuring they don't spiral out of control. For queries related to cybersecurity, biology, chemistry, or distillation, the system fails over to Claude Opus 4.8, a more stable and less powerful model. This failover mechanism is a clever way to ensure that sensitive areas are handled with the utmost care.
The safety measures don't stop there. Anthropic has also updated its data retention policy, a move that's sure to raise eyebrows and spark debates. The company will now retain log data for 30 days for trust and safety purposes. This decision comes amidst heightened scrutiny over how Anthropic manages user telemetry and trust, especially following the discovery of a secret Claude Code tracker that was found monitoring specific user populations. This data retention policy applies even to organizations with zero data retention policies, marking a significant shift in how user data is handled. Anthropic has been quick to assure users that this data won't be used for training new models, but the change is still a notable departure from previous practices.
Pricing: A Hefty Tag for Power
Let's talk about the elephant in the room: the price. Both Claude Fable 5 and Claude Mythos 5 come with a hefty price tag. At $10 per million input tokens and $50 per million output tokens, it's clear that this isn't a tool for casual users. However, it's worth noting that this pricing is significantly lower than the previous Mythos Preview pricing, which was set at $25 per million input tokens and $125 per million output tokens. The reduction in price is a welcome change, but it's still a substantial investment.
For those eager to get their hands on Fable 5, there's some good news. The model is available through Pro, Max, Team, and seat-based Enterprise plans at no extra cost until June 22. After this date, it will require usage credits, so early adopters have a window to explore its capabilities without additional financial burden. Mythos 5, on the other hand, remains exclusively available to Glasswing partners, ensuring that its most advanced features are reserved for a select group.
Performance and Resistance
When it comes to performance, Mythos 5 doesn't disappoint. Early tests have shown strong resistance to prompt injection attacks, a common vulnerability in AI models. With a success rate of just 4.8% over 100 attempts, it's clear that Anthropic has put significant effort into bolstering the model's defenses. This resistance is crucial for maintaining the integrity of the system and ensuring that it can't be easily manipulated for malicious purposes. Bolstered defenses are especially critical given the intense competitive pressures and threats of model replication, such as the massive Alibaba scraping campaign targeting Claude recently reported by Anthropic.
The model's performance in software engineering and knowledge work is particularly impressive. It's designed to handle complex tasks with ease, making it a valuable tool for developers, researchers, and professionals in various fields. The vision capabilities are also noteworthy, opening up new possibilities for AI-assisted image and video analysis. In short, Mythos 5 is a powerhouse, capable of tackling a wide range of challenges with precision and efficiency.
The Road Ahead
The release of the Mythos model is more than just a product launch; it's a statement about the future of AI. Anthropic is pushing the boundaries of what's possible, while also grappling with the ethical and practical challenges that come with such advanced technology. The guardrails and safety measures are a clear indication that the company is taking these challenges seriously, but they also raise important questions about the balance between innovation and control.
As Mythos finds its way into the hands of users, it will be fascinating to see how it's utilized. Will it live up to its potential, or will the guardrails prove to be too restrictive? Only time will tell. One thing is certain, though: the release of Mythos is a significant moment in the AI landscape, one that could shape the direction of the field for years to come.
For now, those interested in exploring the capabilities of Claude Fable 5 can do so through Anthropic's various plans. And for those lucky enough to be Glasswing partners, the full power of Mythos 5 awaits. It's an exciting time for AI, and Mythos is at the forefront of this new era.