Fable and Mythos
Fable 4 is basically Mythos in a straitjacket of safeguards, and the way Anthropic nerfed it reveals their real moat: not the model, but the harness around Claude.
Most model releases are over-sensationalized, but this one is genuinely unique.
In February, Anthropic announced Mythos to the world without releasing it. Instead, they dropped a risk report that basically said the public can't be trusted with it because it lacked the necessary safeguards (essentially, moderation) for blocking high-risk responses, even though they don't technically know the severity of those risks. What they do know is that the model is powerful enough to accelerate its own development.
The release of Fable 4 yesterday is basically Mythos in a straitjacket of safeguards. While it's only been a day, Twitter seems to agree that Fable performs exceptionally across software engineering and knowledge work tasks, as claimed in the Fable system card.
This release is unique because of its dual focus on performance and the omnipresent alignment problem, specifically in the form of recursive self-improvement. These safeguards are particularly sensitive to any question relating to biology or cybersecurity.
Anthropic is covertly degrading their model quality, but only for a specific archetype of user: someone who's trying to use Anthropic's model to build their own frontier LLM. But who actually has the resources to do this? The answer is other frontier labs, who are already rumored to be doing this exact thing.
But it's a little silly, right? Anthropic's real moat isn't in their model's raw capabilities, but in their coding harness for Claude. That's what will actually make customers who find real value in AI stick around in the long run. So by releasing a nerfed Mythos (Fable), they're hitting two birds with one stone:
- Cutting off competitors' ability to copy them, and
- Demonstrating their commitment to safety and compliance on a global stage.
This is timed suspiciously well with the hype around the rumored AI company IPOs. But to Anthropic's credit, so far they've kind of earned their success. They're profitable, they're trusted, and they've managed to avoid every mistake that their main predecessor (OpenAI) has fallen face-first into over the last two years.
I'll be making a follow-up video and post with a TLDR of how Anthropic actually does this, since it's probably the most important principle to understand going forward in this season of AI updates.