Ep. 48: When the Most Powerful Model Comes With a Bouncer
In this episode I cover Anthropic’s latest model release, Fable 5, the publicly available version of their most powerful tech to date. I get into what it is, why it probably isn’t for most people, how the safety features actually work in practice, and the backstory on why the rollout was so carefully controlled in the first place. Worth noting: Fable 5 was taken offline by the US government shortly after this post was written. The information provide still stands however, and honestly, the shutdown just adds another layer to the story.
Left Off the Bat: This Model Probably Isn’t for You
Today we are talking about Anthropic’s latest model release: Fable 5, and right off the bat, two full disclosures:
- Two days after I wrote this post, Fable 5 was shutdown. So what you’re going to read below still applies because it’s just explaining what Fable 5 was, but you can’t go and try it for yourself.
- Prior to the shutdown, I was only able to use Fable 5 to outline this episode, which is honestly not what it’s even intended for, but I wanted to at least try it before talking about it. So, my experience isn’t coming from some deep usage or anything like that.
Left off the bat, this model probably isn’t for you. Which means that it being shutdown probably doesn’t affect you much, but it’s still worth being aware of the model IMO.
I have no idea who everyone is who listens to the podcast or reads this newsletter, but with how expensive Fable 5 is and what it’s really built for, it is very likely not for you (or me!).
But we’re curious people here, and I like keeping y’all informed. So I figured I’d do a brief episode on Fable 5 so you are at least aware of it, have some information about it, and can then go do with it what you’d like.
The Shutdown
Fable 5, the newest model in Anthropic’s line of models was released to the public on June 9th. Just three days later, on June 12th, Anthropic received a government issued legal directive ordering it to cut off all foreign nationals (including its own foreign-national staff). Faced with an impossible task, Anthropic complied by disabling Fable 5 and Mythos 5 worldwide.
The government’s stated concern was that someone had found a way to get around Fable 5’s safety guardrails. Anthropic reviewed the claim and found it to be a narrow workaround, not a broad bypass of the model’s protections, and noted that other publicly available models (GPT-5.5) have the same vulnerability. Honestly, it reads like retaliation to me (remember when Anthropic refused to let the government have unrestricted access to Claude for mass domestic surveillance and fully autonomous weapons systems?).
Anthropic pushed back publicly but did comply with the order, and as of today, both models are still dark, while Anthropic reports they are “working to restore” access.
The 30,000-Foot View
Fable 5 is the first publicly available version of Anthropic’s Mythos model, which is their most powerful tech to date. I spoke about Mythos a few episodes back when it was causing a stir in the news and had the cybersecurity world in a tizzy.
Think of Fable as the new it girl. The lineup used to top out at Opus, but the Mythos-class now sits above that.
And yes, we have two names here: Mythos and Fable. So, name decoder time:
- Mythos 5 is the unrestricted version for vetted partners
- Fable 5 is the same model with safeguards applied, and is what was released to the public
- The 5 is because they are Anthropic’s 5th generation models
The Backstory
Anthropic originally unveiled the first version of Mythos in April, and limited the rollout because of its advanced cybersecurity capabilities. Shit was just too powerful.
As a countermeasure, Anthropic actually rolled out Project Glasswing at the same time, and gave select organizations early access to Mythos. These organizations were to use Mythos to find and fix software vulnerabilities in critical infrastructure (think financial institutions, software companies, and healthcare networks) before the wide release of Mythos.
Basically Anthropic was trying to have the folks of Project Glasswing use Mythos to patch any security vulnerabilities, so that the “bad guys” couldn’t use Mythos to exploit those vulnerabilities.
Apparently just not releasing it at all was not an option.
So How Is It Safe to Release Now?
So glad you asked.
Anthropic released Fable 5 with guardrails.
Fable 5 uses classifiers that flag high-risk requests in three areas: cybersecurity, biology/chemistry, and distillation. For those of you like me, who have no idea what distillation is, Distillation is when a company trains a cheaper, smaller model by feeding it the outputs of a more powerful one. The small model learns to mimic the big model’s responses without needing the big model’s compute or architecture. It’s how you get “80% of the capability for 10% of the cost.
Translation, you can’t use Fable 5 to build a cheaper LLM.
The guardrail is that any flagged requests (cybersecurity, biology/chemistry, and distillation) automatically get routed to the older, less capable Opus 4.8 model. Apparently over 95 percent of sessions are unaffected, but of course Threads users were quick to take to their keyboards and complain about tasks that were wrongfully re-routed.
Plain English: ask it something sketchy and a different, less capable model answers instead. Enter the bouncer analogy I used for the title of this episode. Yes, it’s the most powerful model in the room, but it comes with a bouncer at the door.
What It’s Actually Good At
The headline for this model release isn’t “smarter answers to your questions.”; the headline is stamina.
Fable 5 can work autonomously for longer than any prior Claude model and outperforms Opus on longer, more complex tasks.
Strong areas:
- Coding and software engineering
- Long-running autonomous tasks
- Scientific research
- Financial analysis
- Legal document work
Who Actually Needs This
So who is Fable 5 actually built for? If you’re vibe coding big projects, running long agentic workflows, doing deep multi-step research: this is for you.
If you’re drafting emails, brainstorming, asking questions: stick to the everyday models. My driver continues to be Sonnet 4.6.
And once again, we see AI with their backwards marketing as it relates to the masses: “Here’s a solution, go find a problem.” I’m not currently building anything or doing any vibe coding, so I literally had no problem for it to solve, but perhaps you’re in a different boat.
Access and Money
Fun fact: Fable 5 is basically just available as a taste test right now.
It’s included in Pro, Max, Team, and seat-based Enterprise plans now, but only through June 22. On June 23, it gets pulled and requires usage credits, aka additional monies on top of whatever you’re paying monthly for Claude.
Translation: there’s a free trial window happening right now. If you’re curious, go play with it before the 22nd.
Worth noting: this model is token HUNGRY. It draws down usage 2x as fast as Opus, the next most capable model, which was already notoriously expensive. So, use it, but keep an eye on that usage bar.
But What About…?
Something worth noting: This launch came only days after Anthropic publicly urged major AI labs to establish a coordinated brake pedal on frontier development, and all while they’re prepping for an IPO. Insert facepalm.
We both know that brake pedal 100% isn’t going to happen.
Are we headed towards Skynet? No doomerism here, so I’m gonna leave that answer up to you.
How I Used AI This Week
Each episode I share a quick example of how I used AI that week. This week we’re keeping it meta, because my use case for AI was asking Fable 5 to explain itself and then help me with the outline for this podcast episode.
It honestly felt zero percent different than when I’ve used Claude in the past to help with podcast episode research and outlines, but I suppose that’s just a part of being curious.
Da Wrap-up
Fable 5 was Anthropic’s most powerful model to date, but I have no idea when it’ll be coming back. Perhaps this is ultimately for the best, but something tells me probably not. Cheers to the Terminator.
As always, endlessly grateful for you and your curiosity.
Catch you next Thursday.
Maestro out.