.svg)
.svg)
For Instructor
Greetings, AI enthusiasts! If you thought that AI was progressing quickly prior to this point – get ready to be blown away! Anthropic has come out with a product called “Mythos 5” that will reframe how we think of Generative AI. This is their most advanced model, pushing the limits of technology forward by light years.
Typically, when there is a new AI model released, the tech community will spend a few days going nuts talking about benchmarks (chart comparison, math score comparisons, etc.) and arguing on social media about whether the new model can generate code in Python a few percentage-points better than the last model. However, with Mythos 5 we are throwing all traditional benchmarks out the window because Mythos 5 currently beats all existing models across a huge number of benchmarks including their own Mythos preview model released only a few months ago.
With that said, there is virtually no value in talking about benchmark results today. The achievements Mythos 5 has made with real-world, highly complex applications is far more important than a standard set of sterilized benchmark performance scores. We are witnessing a huge jump in the capabilities of this model.
The Autonomous Scientist: The lab use of Mythos 5
To Folio 55 (the first company to develop artificial intelligence laboratory helpers), the best way to comprehend the degree of how exceptionally innovative Mythos 5 is would be to examine how it currently operates within the biological/healthcare setting. My conclusion is based on my analysis of several evaluations that demonstrated an autonomous exposure to Mythos 5 conducting complex active drug discovery (actively searching for, identifying actual new drug candidates) rather than simply providing assistance to an individual scientist for writing a report or organizing scientific research. In fact, using an objective analysis methodology to identify 9 out of 14 different disease areas where Mythos 5 produced a candidate drug without the assistance of a human scientist has validated this concept.
Consider for a moment that the process of introducing any new pharmaceutical compound to the world takes multiple years of iterative testing, trial-and-error experimentation involving thousands of highly skilled scientists working together, and billions of dollars in funding. Imagine navigating this extraordinarily complex environment of chemistry and biology to identify potentially life-saving drug treatments completely unassisted by any human individuals. The implications for both the medical/scientific community and the global pharmaceutical industry have been nothing short of revolutionary.
The Twist to Your Dreams
You might be dying to get Mythos 5 for your own vprojects, but sorry, there are some catches to it: there are no applications yet. Up to this point, Mythos 5 can only be obtained through Project Glasswing, which is an extremely selective initiative designed specifically for organizations that qualify and meet the high standards required by Anthropic to receive an AI of Mind-Blowing Capability. Anthropic has purposely kept tight security around the Mythos AI ecosystem because of the complexities involved as AIs become increasingly intelligent and the associated rise in risks associated with having access to AIs with such intelligence.
Announcing Fable 5: Your "Nerfed" Superpower
Anthropic has also given us something to work with besides Mythos 5. We get an inferior but still highly capable version of Mythos 5 called Fable 5. Now when I say "inferior" that does not mean weak; Fable 5 and Mythos 5 are built on the exact same systems and architectures; they differ only in that Fable 5 comes with some very robust safety guard rails integrated into the product to prevent individuals from using them to develop bioweapon technologies or conducting complex cyber attacks.
Despite these significant limitations, Fable 5 is doing things never before possible.
Real-World Example of Fable 5: Migrating Millions of Lines of Code
Stripe, a world-class payment processing provider, was one of the first companies to have access to Fable 5. They put Fable 5 through an Ultimate Test by using the product to migrate over 50 million lines of legacy code in Stripe's backend systems.
If a team of skilled human engineers were to attempt this monumental task, it would take them at least two months of intense, tedious, and highly coordinated work to complete. Fable 5 completed the entire migration in just one day! One Day! It determined how it would migrate the code, performed the code changes, validated the changes against its own standards, and moved on! The effects for software engineers and enterprise IT are truly mind-blowing - you’re talking about months of engineering work being done in just a few hours!
The Game of Pokémon and the Challenge of Financial Mastery
If that sounds boring, let’s talk about video games! We have seen AI play games previously, but the most recent, Fable 5, has played Pokémon FireRed solely using raw screenshots of the game, without access to the underlying code, API, or memory states; it looked at the screen the same as a human would, interpreted visual data, and then utilized its advanced vision abilities to create strategies to win the game.
But not only does Fable 5 excel at playing video games, it also ranks higher than any model ever tested on Hebia’s rigorous finance benchmarking, demonstrating that it can support complex, structured reasoning, and perform sophisticated financial analysis without breaking a sweat.
The Shift in Perspective: From Activity-Oriented to Responsibility-Oriented
The most significant point of all is clearly articulated by Felix, head of Claude Code and co-working on Desktop at Anthropic; we are in a new era in Computing.
All of the activity in Generative AI for the last few years has been about assigning AI activities/tasks to perform. An example of this is asking an AI system to write a polite email, summarize a lengthy PDF document, or provide a short example of boilerplate code. Once it completes the activity, an AI system will return the result and passively await the next instruction.
With frontier models such as Mythos 5 and Fable 5, we are moving rapidly toward moving from assigning AI activities to assigning AI responsibilities.
Asking an AI system to "create a Python script" represents assigning a minor (insignificant) activity to the AI system. Asking an AI system to "safely move our 50 million line production code base by tomorrow morning" represents significant (major) responsibility. Asking an AI system to "research and identify prospective drug candidates for this rare disease" represents life-altering (huge) responsibility.
Solving this issue is a far larger challenge than others we’ve faced. If you’re designing AI to handle large projects, which run continuously and permit considerable time to complete, including intelligent capability to plan, self-correct errors, verify work, and respond to unanticipated obstacles without needing to contact a human operator for assistance, it may be a major challenge.
Anthropic's Cautiousness
Due to the dramatic increase in functionality of the latest models, Anthropic has taken a very cautious attitude towards releasing these models into public use. Relying heavily on AI for such large-scale tasks introduces the possibility of catastrophic failure if something goes wrong, particularly when it involves real-world scenarios. If Fable 5 makes an error in drafting an email, the worst thing that will happen is you may embarrass yourself in front of your boss. However, if an unrestricted Mythos 5 is released into the hands of malicious users, so that the mythos 5 can be asked to locate zero-day vulnerabilities on the nation’s electric grid; then the consequences of such a mistake will have enormous ramifications.
Therefore, aside from the tremendous responsibility; Fable 5 has been engineered with a brand new set of sophisticated “safety classifiers.” A request for highly sensitive information, such as how to break into a computer system for offensive cybersecurity purposes or break through biosecurity (to obtain drugs or human tissue), is intercept by a completely separate AI system, which blocks your request and sends you to the previous version of it (Opus 4.8). Anthropic needs to maintain a very fine balance between granting us access to a powerful autonomous agent, while doing everything in their power to maintain the guards around the new technology.
