The Rise And Fall Of Fable 5

WHAT IT ACTUALLY WAS

Mythos 5 is the underlying frontier model. Fable 5 was described as a Mythos-class model made safe for general use. Same model, safeguards bolted on. The important word is Mythos-class, which sits above the Opus class. This was the top of the Claude 5 generation, and until now Mythos had been locked to roughly 200 organizations, including US government, under something called Project Glasswing. Fable 5 was the first time a Mythos-class model went public at all.

The benchmarks were not a modest bump. It posted the best scores anyone had recorded on nearly every test it was put through, more than 10% over Opus 4.8 on some tasks, strongest reported result on Cognition’s FrontierBench for coding, top of the Hebbia finance benchmark, and the first model to break 90% on core analytics on Hex. The vision side could reportedly rebuild a web app’s source from screenshots. It held coherence across millions of tokens of long-horizon work. Pricing was 10 dollars per million input, 50 per million output, free across the paid tiers for two weeks, also on Bedrock.

So this was real. I want to say that plainly before I get to the fall, because the fall does not make sense unless you accept that the capability was genuine.

THE RECEPTION

If you were anywhere near the feeds that week, you saw it. Overwhelmingly enthusiastic, with the usual second thread of slow, expensive, over-guarded. The big names lined up to call it a step change, the best model out there by a clear margin, something of a beast. The line I think would have caught our engineers is the one about long-horizon work: apps that took a hundred prompts a year ago now getting oneshotted, whole classes of problems that used to be out of reach suddenly in reach. That is the kind of claim that makes you want to stop reading threads and throw a real ticket at it.

The build claims were large and mostly vendor-reported, so discount accordingly. The one that matters for a shop like ours is Stripe reportedly compressing a 50-million-line migration into a single day instead of two months. That is squarely the kind of work we sell. If even half of it holds up, that is the headline for us, not the benchmark table. The rest, oneshot apps, UI and game code, the 90% analytics number, is the same story pointed at other work.

And the irony was not lost on anyone. Anthropic shipped its most powerful public model days after warning that AI is getting too dangerous. That tension is the whole story, really, and it is the part I would have wanted us talking about before we got excited about the migration demo.

THE FALL

Then a few days later it was just gone. The short version: a US exportcontrol directive landed citing national security, and Anthropic was reportedly given about 90 minutes to comply. The order was to cut off all foreign nationals, including Anthropic’s own foreign-national staff. Sorting access at that level on that clock was apparently impractical, so they took the blunt path and switched off both Fable 5 and Mythos 5 worldwide, for everyone. Other Claude models stayed up. Sit with that for a second from our seat. This is not a model that got nerfed or ratelimited. It is a model that was working on Thursday and gone by Friday because of a phone call none of its users were part of.

The stated reason is security, and the specific mechanism is almost mundane: a jailbreak where you ask the model to read a specific codebase and fix any flaws, which could surface information useful for cyberattacks. Mythos-class cyber capability is the part that makes this serious. Autonomous zero-day discovery at scale, a reported 83.1% on a vulnerability-reproduction benchmark, though that number is secondary and press-reported, so hold it loosely. There is also a single secondary report that Pliny the Liberator published Fable 5’s system prompt within about 48 hours.

Anthropic pushed back hard. They said the jailbreak only surfaced a small number of previously known, minor vulnerabilities, that similar capability is already widely available from other models, and they named GPT-5.5. They called the recall an overreaction, said they are working to restore access as soon as possible, and sent staff to DC for a meeting set for June 22.

THE REVERSE – ENGINEERING PROBLEM

Here is the angle that got less coverage. Semafor reported on June 13 that the export-control decision was partly motivated by suspicions that a China-linked group had accessed Mythos. I want to be clear about how solid this is: it is suspicion, single-source, not a confirmed breach. The fear underneath it is knowledge distillation, where you train a student model to imitate a target’s capabilities without ever touching the weights, and in doing so replicate the offensive-cyber ability.

There is also a contradiction worth flagging. David Sacks reportedly said the White House learned Fable 5 could be jailbroken, claimed Amodei called it not a serious risk, and said Anthropic refused to fix it before controls, calling the whole thing serious but easily resolved. But an Anthropic spokesperson said the White House did not raise Chinese access in the jailbreak talks at all. So that thread is not just unconfirmed, it is internally contested. Treat it as noise that might be signal, not as fact.

And honestly, I hold two things at once here. Part of me thinks reverseengineering would not be the worst outcome. If a Mythos-class capability gets distilled and re-emerges, it most likely shows up as an open-source model, and that broadens access rather than concentrating it in 200 vetted orgs. I am not reflexively against that. But I also concede the security concern is genuine, and it is genuine precisely because of the capability. A model that can autonomously find and exploit vulnerabilities at scale is not a normal thing to want loose in the world. Both reads are true. I am not going to pretend one cancels the other.

But the more I sit with it, the more I think the espionage worry is only part of the reason this got pulled, and not even the biggest part. Strip it out entirely and you are still left with the plain version of the problem. For three days, anyone with a credit card could rent a Mythos-class model. That includes bad actors. A model that can find and exploit vulnerabilities on its own does not care whose hands it is in, and a public release puts it in everyone’s hands at the same time. You do not need an espionage angle for that to be dangerous. You just need the model to exist and be reachable.

That is the part I keep landing on, and it is where I move off the fence. At this level of capability, the risk of pointing it at the whole internet starts to outweigh whatever we gain from open access. I do not say that lightly, because I am usually the one arguing for more access, not less. But Fable 5 is the first model where I think the math actually flips. Pulling it back to vetted, supervised use is not the instinct I want to have. It is just the one I cannot argue my way out of here.

Where this is heading, if it comes back at all, looks like vetted access. The vetted-corporations-and-government-partners framing is really a description of Glasswing, the pre-launch arrangement, but it is also the most likely shape of any return. US-nationals-only, narrow, supervised. Not the open two-week free trial we got for three days.

WHAT THIS MEANS FOR US

I keep coming back to the same lesson from issue eight. The model release itself is now part of the risk surface. Fable 5 is the cleanest example yet. This was not a model that degraded or got nerfed. It was a model that was capable enough to become a national-security object and got switched off by directive in 90 minutes. You cannot benchmark your way around that. No amount of capability protects a workflow built on a single model that a government can pull on a Friday afternoon.

Last issue that worry was still mostly theoretical. A bad release could quietly degrade a working system. Fable 5 makes it concrete. Any model we do not fully control is now a genuine disruptor to an established workflow, not someday but already. The provider can nerf it, price it out, gate it, or have it pulled out from under us by a government overnight, and not one of those is a thing we get a vote on.

So what this means for us is simple, and it is the thing we have been pushing on for the past few weeks. A model-agnostic workflow is no longer a nice-to-have. It is the priority. This is the core principle I want us building everything around from here. The workflow should not care which model sits underneath it. The model becomes an interchangeable part, and we keep the ability to test, compare, switch, and recover without rebuilding the whole setup every time the ground moves under us. The more we bake that principle in now, the less any single launch, recall, or price change can hurt us later.

That is also our best defense against enshittification. When a model we lean on gets worse, more restricted, or more expensive, and at some point one of them will, a model-agnostic setup is what lets us swap to a better alternative the same day instead of being held hostage to one provider’s decisions. It is a powerful backup precisely because it is built in advance, not scrambled together after the thing we depend on breaks. We were never going to build on Fable 5 in its three days of life, but plenty of teams reportedly tried, and they got a very fast lesson in why you keep a second option warm.

The capability was real. That was never the question. The question is whether the thing you are building on will still be there next week, and Fable 5 just gave us the most vivid possible answer.

Sources → Anthropic launch · Anthropic on the suspension · TechCrunch · CNBC · VentureBeat · The Conversation · Semafor (China angle) · Tom's Hardware (Sacks)