Fable 5's return

The most capable AI model ever released was shut down by the US government then returned, a wild saga with vital lessons for AI policy for schools.

In this episode:

• Anthropic's Claude Fable 5, the "most capable AI model ever released," experienced a rapid launch, government-mandated shutdown, and return, offering a blueprint for future AI model reliability challenges.
• The 22-day saga of Claude Fable 5 underscores that a robust AI policy for schools must include contingency planning to avoid a single point of failure when integrating AI into workflows.
• The US government's swift intervention with emergency export controls on Fable 5 signals that government AI regulation is a dynamic landscape, necessitating adaptable AI in education strategy based on principles, not specific tools.
• Fable 5's comeback, facilitated by enhanced safety classifiers and a HackerOne program, illustrates that transparent, iterative improvement in AI safety is achievable and builds trust.
• Schools should model Anthropic's transparency by communicating openly about AI tool limitations and actively working to improve them, fostering reflective awareness of AI's capabilities and constraints.

Chapters:

• 00:00 — Cold open & welcome
• 00:30 — The wild saga of Claude Fable 5's launch and shutdown
• 01:25 — Understanding Anthropic's Mythos-class models and safeguards
• 02:40 — Staggering early performance and pricing of Fable 5
• 03:20 — Why the US government pulled the plug on Fable 5
• 04:35 — The engineering and diplomacy behind Fable 5's return
• 05:40 — Lesson 1: Contingency in AI policy for schools – avoiding single points of failure
• 07:15 — Lesson 2: Adapting to government AI regulation with principle-based policies
• 08:25 — Lesson 3: The hopeful truth about AI model reliability and transparency
• 09:40 — The enduring task for educators in an age of extraordinary AI capability

What is Anthropic Fable 5 and why was it temporarily shut down?
Anthropic Fable 5 was a highly capable AI model launched by Anthropic that was temporarily suspended by the US government due to national security concerns after researchers bypassed its safeguards.

How does the Fable 5 incident inform AI policy for schools?
The Fable 5 incident teaches schools to build contingency plans for AI tools, assume dynamic government AI regulation, and anchor their AI in education strategy in adaptable principles rather than specific products.

What does the Fable 5 story tell us about AI model reliability and safety?
Fable 5's return demonstrates that AI model reliability can be enhanced through dedicated engineering work on safeguards, transparent collaboration with regulators, and open communication about limitations, proving safety and capability are not mutually exclusive.

Featuring: Dan Fitzpatrick, Anthropic, Claude Fable 5, Project Glasswing, Claude Opus 4.8, Mythos 5, US government, HackerOne, OpenAI GPT-5.5.

Follow AI in Education with Dan Fitzpatrick for more on AI in education.

Transcript

SPEAKER_00 0:00

If this episode makes you think, please let us know in the comments and support us by subscribing and leaving a review. Thank you. Today we are exploring the Return of Fable Vhat Three Weeks of AI Whiplash Should Teach Every School Leader An article by me, Dan Fitzpatrick, published in Forbes on july second, twenty twenty six. This piece unpacks a truly wild episode in the short history of Frontier AI, and it holds some really potent lessons for everyone in education. The most capable AI model ever made available to the public was launched on June 9th, then effectively switched off three days later by the US government, only to make its comeback on July 1st. That's a 22-day saga that frankly you couldn't make up, and it's a profound teaching moment for anyone navigating AI in our schools. So let's break down what actually happened here, because it sets the stage for everything else. This story revolves around a model called Claude Fable V from Anthropic. The company launched it on June 9th, positioning it as their first mythos class model for the general public. According to Anthropic's launch announcement, Fable V was state of the art across pretty much every benchmark you could think of. It was showing exceptional performance in areas like software engineering, knowledge work, even scientific research. What Anthropic highlighted was that the longer and more complex the task, the more Fable V pulled ahead of its predecessors. Now there's some important backstory here. Back in April, Anthropic had actually unveiled a model called Mythos. This thing was so ridiculously good at finding security flaws in software that the company initially refused to release it publicly. Instead, they restricted it to a small handpicked group of cyber defenders through an initiative they called Project Glasswing. Fable 5, the model we're talking about, is essentially that same underlying Mythos model, but anthropic wrapped it up in the strongest safeguards they had ever built. Its close sibling, Mythos 5, keeps some of those safeguards lifted in specific areas and remains exclusively available to approved organizations working directly with the US government. These safeguards on Fable V were quite unusual by design. What would happen is if Fable V detected a request in a really high risk area, think cybersecurity or biology, those kinds of fields, it would automatically hand that query over to Anthropic's next most capable model, Claude Opus 4.8, instead of processing it itself. TechCrunch reported that during early testing, at least 95% of user sessions ran entirely on Fable V's own responses, and Anthropic's pre-launch red teaming, which is basically trying to break the model, didn't produce any universal jailbreaks across more than a thousand hours of testing. The early performance stories coming out of this model were just staggering. Anthropic itself reported that a company called Stripe used Fable V to complete a codebase-wide migration across a massive 50 million line Ruby codebase in a single day. That's work that was estimated to take a full engineering team more than two months. Another analytics firm, Hex, said Fable 5, was the very first model to score 90% on its core benchmark of complex, long-running analytical tasks, according to TechCrunch. You can imagine the excitement, but all this power didn't come cheap, as you might expect. The pricing landed at $10 per million input tokens and $50 per million output tokens, which was double the price of Claudopus 4.8, though still less than half the cost of the earlier super restricted Mythos preview. Then, just three days after this incredible launch, the government pulled the plug. On June 12th, less than a week after Fable Vent public, the US government applied emergency export controls to both Fable V and Mythos V. They cited national security concerns, and the order essentially required Anthropic to restrict access for foreign nationals, regardless of whether they were inside or outside the United States. The problem was Anthropic simply had no reliable way to verify nationality in real time, so they had to make a tough decision. They suspended access to both models for everyone, globally, immediately. Can you imagine the whiplash? The trigger for this, as reported by Moby Health News and others, was a report from Amazon researchers who had managed to bypass Fable V safeguards. In one specific case, the model was prompted to identify a software vulnerability, and then it went further, producing actual code demonstrating how that vulnerability could be exploited. When that report landed on government officials' desks, alarm bells naturally followed. Anthropic pushed back hard, I'm sure. The company argued that the exploit didn't really tap into anything uniquely dangerous. They pointed out that their own testing had shown other lead-in models, including Claude Opus 4.8, OpenAI's GPT 5.5, and even Moonshot's Kimi K 2.7 could identify the exact same vulnerabilities and produce the same kind of demonstration. Some independent security experts even called the jailbreak concern significantly overblown. But the government evidently was completely unmoved by the everyone else can do it too defense. And honestly, I have some sympathy with that position. When it comes to national security, you can see why they'd err on the side of caution. So how did Fable V manage to come back? It was a really fascinating mix of some serious engineering work and a whole lot of diplomacy. Anthropic developed and trained an improved safety classifier, specifically designed to catch and neutralize the exact technique the Amazon researchers had used. They also launched a new Hacker One program, inviting security researchers from all over the world to submit potential cyber jailbreaks for review, and they proposed a new industry-wide framework for how to handle these kinds of jailbreak discoveries. Behind the scenes, the company spent ten intense weeks working with various government agencies, the Office of the National Cyber Director, the Office of Science and Technology Policy, the Treasury, and the Commerce Department. All of this collaboration was happening as the administration was developing its June executive order on AI innovation and security. There was a political dimension to this too. Wired reported via VentureBeat that Anthropic's initial argument that no frontier lab could guarantee zero jailbreaks really frustrated officials. They reported that real progress only started to happen when the company shifted its focus from debating the theoretical impossibility of perfect safety to actually demonstrating stronger safeguards in practice, and it worked. On June 30th, the export controls were officially lifted. Anthropic announced that Fable 5 would return globally from July 1st across all its platforms. The Clawed Platform, Clawed AI, Clawed Code, and Claude Cowork. Mythos 5 access was also restored for approved US organizations following government sign-off on June 26th. For those who are subscribers, it's worth knowing that Pro, Max, Team, and Select Enterprise Plans could use Fable 5 for up to 50% of their weekly usage limits through July 7th, after which access moves to usage credits. Cloud availability on AWS, Google Cloud, and Microsoft Foundry is being re-enabled as quickly as possible. Now I can hear the objection already, right? Dan, this is a cyber security story about some super frontier model that most teachers will never consciously choose to use. Why on earth should a head teacher in Carlisle or a superintendent in Connecticut care about any of this? Well I think there are three absolutely critical reasons why this matters to every single one of you in education. First, the dependence lesson. Think about this. For nineteen days, individuals and entire organizations who had built their workflows around what was, at that point, the most capable AI model on the market simply couldn't use it. One developer posted on the McRumers forums described spending five hours with the restored Fable Vixing three weeks of accumulated errors from having to use a lesser model. This is huge. Schools are increasingly building AI into their daily operations, into planning, into assessment, into administration. If your entire provision, or even just a critical part of it, depends on one model from one company in one regulatory jurisdiction, then you don't have an AI in education strategy. What you actually have is a single point of failure. This Fable 5 saga is a stark reminder that when you're thinking about AI policy for schools, you simply must build contingency into your AI plans just as you would for any other critical system in your school. Think about a year eight history teacher who might rely on a specific AI to quickly generate diverse historical scenarios for a classroom debate. What happens if that tool vanishes overnight? We have to proactively consider these disruptions. We're talking about outsourcing your doing, not your thinking, but we can't outsource our operational resilience. Second, the governance lesson. The Fable V episode is the clearest signal yet that governments and particularly the US government will intervene in AI deployment at speed, without warning, and with blunt instruments. Forbes reported that these restrictions were only lifted after anthropic satisfied the administration's specific operational concerns. What this means for school leaders writing Org Clechar Faris AI policy for schools is that you should absolutely assume the regulatory ground will keep moving. You need to write policies that are anchored in principles and processes, not in specific products or specific versions of tools. This really reinforces that start with why not how philosophy. A department head planning CPD on AI tools, for example, needs to emphasize critical evaluation and adaptability above all else, rather than just teaching staff how to use a specific feature of Tool X, because Tool X might not be around next year, or its capabilities might be severely curtailed by government AI regulation. If you find these insights valuable and want to stay ahead in understanding AI's impact on education, make sure to follow and subscribe to the podcast wherever you listen. Third, the safeguards lesson, and this one I think is a genuinely hopeful one. Fable five's return demonstrates something I've argued for years. Safety and capability are not actually opposites. Anthropic deliberately tuned its classifiers to be cautious, openly acknowledging in their announcements that benign requests would sometimes be blocked and that false positives would yes, frustrate users. The honesty, that transparency is the real point here. A company that's willing to say our guardrails are imperfect, here's how we're improving them, is modelling exactly the kind of posture schools should take with their own AI use. Transparent, cautious, and iterating in the open. This kind of approach is essential for building AI model reliability and fostering trust. Think about a school piloting a new AI tool for differentiation in a year five literacy class. Instead of expecting perfection from day one, they could communicate openly with parents and students that they are testing it, learning its limitations, and actively looking for any potential bias or missteps, much like Anthropic did with their Hacker One program. It's about developing that reflective awareness and understanding AI limitations. So the bottom line is this the most capable AI model ever released to the public went from launch to lockdown to comeback in just 22 days. This isn't just some footnote and tech news, it's a stark preview of the decade ahead, a decade where extraordinary capability, genuine risk, and swift government intervention are going to keep colliding. For us as educators, the fundamental task hasn't changed. We need to stay curious rather than fearful. We need to build for evolution, not revolution. And we must always remember that the thinking, the judgment about how these tools genuinely serve learners, how they deepen understanding, how they foster curiosity and care, that absolutely can never be outsourced no matter how capable the machine doing the doing becomes. Extraordinary capability, genuine risk, and government intervention are the constants. Our task as educators is to remain curious, build for evolution, and protect the human judgment that no machine can replicate. That's all for today. Thanks for listening.