AI in Education: Preparing Students for Mythos-Class AI

Anthropic's new Mythos-class AI, Claude Fable 5, compressed two months of human work into a single day for Stripe. This changes everything for AI in education.

In this episode:

• Anthropic's new Mythos-class AI, Claude Fable 5, achieved a 50-million-line codebase migration for Stripe in one day, a task estimated to take humans two months, signifying a major leap for AI in education.
• Effective teaching with AI requires fostering 'task imagination' in students, enabling them to define multi-day projects for AI and articulate clear quality criteria.
• AI assessment for educators should evolve to evaluate students' ability to direct and critically judge AI-generated work, rather than just their capacity to perform tasks themselves.
• Strict safety classifiers on Claude Fable 5, sometimes rerouting science queries, provide valuable, live examples for teaching AI literacy in schools about governance, ethics, and the dual-use dilemma.
• School leaders deploying AI for school operations must carefully examine usage-based pricing models for new AIs like Claude Fable 5 and review data retention policies (e.g., 30-day retention) against data protection obligations.

Chapters:

• 00:00 — Cold open & welcome
• 00:30 — Introducing Claude Fable 5: A Mythos-class AI and its impact on education
• 01:25 — Beyond benchmarks: Fable 5's leap in delegation and responsibility
• 02:30 — The missing skill: Preparing students for 'task imagination' with AI
• 03:45 — Real-world AI literacy: Dual-use dilemma and Fable 5's safety guardrails
• 05:00 — Teaching with AI: Ethics, judgment, and critical thinking with Fable 5
• 06:00 — Nuances for school leaders: Pricing and data retention for AI in education
• 07:30 — The future of AI assessment: Directing and judging work, not just doing it

What is Mythos-class AI and how does it change AI in education?
Mythos-class AI, exemplified by Anthropic's Claude Fable 5, can autonomously manage complex, multi-day projects, requiring educators to prepare students to 'delegate well' and develop 'task imagination' rather than just perform tasks themselves.

How can teachers use AI marking safely with advanced models like Fable 5?
While Fable 5's primary use isn't marking, its underlying principle of delegating responsibilities rather than discrete tasks means teachers should focus on designing comprehensive AI assessment for educators that evaluates students' ability to direct and judge AI work, while remaining vigilant about data retention policies.

What is 'task imagination' and why is it important for AI literacy in schools?
Task imagination is the ability to define a large, multi-day project for an AI, articulate precise quality criteria, and then evaluate its output; this skill is crucial for AI literacy in schools as advanced AIs like Claude Fable 5 demand clear, complex briefs to operate effectively.

Featuring: Dan Fitzpatrick, Anthropic, Claude Fable 5, Opus, Mythos-class, FrontierCode, Stripe, Felix Ryberg, Nate B. Jones.

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 a truly significant development from the world of AI, the launch of Anthropics Claude Fable V, a new mythos class model that would, according to early reports, helped Stripe complete a massive migration across a 50 million line code base in a single day. Work that company estimated would have taken an engineering team over two months. Now that's a number that makes you sit up, isn't it? This wasn't just another incremental upgrade. What we saw this week was a genuine leap forward, and it's already sparking conversations that go way beyond just technical benchmarks. Most of the early buzz has focused on what this model, Fable V, can do. But for us in education, the more important question, the one that really keeps me thinking, is what it changes about the humans who are working alongside it. Because beneath all the impressive statistics and the occasional controversy, Fable V represents a fundamental shift in the relationship between people and artificial intelligence, and that shift lands squarely on education's doorstep. This isn't just about integrating AI in education, it's about reimagining how we prepare students for a world where AI is a deeply embedded partner. First, let's just briefly cover the capability story because it really is something. Anthropic announced that Fable V and its slightly more restricted sibling, Mythos Vare the same underlying architecture. Fable is the version that's been wrapped in all the important safety classifiers for general release. What's truly astonishing is its performance on new benchmarks. For example, on Cognition's new Frontier Code benchmark, which isn't just about whether AI written code passes tests, but whether it's actually good enough for a human maintainer to merge into a project. Fable 5 more than doubled the score of Anthropic's own Opus 4.8. And Opus 4.8 was only released weeks earlier, mind you. That's rapid, isn't it? The speed of evolution here is incredible, not a revolution, but a blindingly fast evolution. But that stripe example, that's the one that really puts things into perspective for me. A 50 million line code base migration, two months of human work compressed into a single day by the model. The number is impressive, absolutely. But what matters more, the pattern behind it, is something we all need to pay attention to. The early users who are getting the most out of Claude Fable V aren't asking it simple questions or even just handing it discrete tasks. No, they're handing it responsibilities. Felix Ryberg, who leads the Claude Code and Cowork Initiatives at Anthropic, described this shift in a public post recently. He used to ask the model to investigate a particular crash report. Now he's got it running a continuous loop, watching every crash report that comes in. Its job isn't to help fix a crash, it's to help keep the apps from crashing in the first place. That distinction, moving from a specific task to an ongoing responsibility, sounds subtle, doesn't it? But it's not. It's the difference between having an assistant and having a genuine colleague. And it raises a question that honestly I think schools have barely even begun to ask. What does it mean to prepare students to delegate well? This connects directly to our core philosophy of human in the loop, an outsourcing the doing, not the thinking. The human is still in charge, still holds the complexity, but the nature of that charge is fundamentally shifting. Here's the uncomfortable truth this launch exposed. We now have publicly available artificial intelligence that can work autonomously for hours, sometimes an entire working day on a single overarching goal. Anthropic's announcement and those early user reports describe runs that stretch across an entire workday, and most of us, educators, absolutely included, have no idea what kind of work to give it. We're great at breaking things down into small tasks, but imagining something big, something that takes days, that's new territory. AI strategist Nate B. Jones, in a fantastic video he put out this week, put a name to this missing skill. He called it task imagination. His observation was that historically we humans have really operated in two main modes when it comes to delegating. We either wave our hands and give a really vague guideline, hoping for the best, or we just do all the detailed work ourselves because it feels easier. But working with models like Fable V demands a third mode. It requires you to articulate exactly what you want, what the bar for quality actually is, and precisely how success will be judged. And then crucially you have to step back and let the AI run. As Jones pointed out, most people simply cannot currently name a single piece of work they could hand to an AI that would take days to complete. And I've seen this firsthand in my training sessions with school leaders. When I ask them to describe a multi-hour task or even a multi-day project that an AI could genuinely take on for their school or their department, the conversation often stalls. It's not about the AI's capability, it's about our capacity to even imagine that brief. We're conditioned to micromanage or to just do it ourselves. The limiting factor isn't the tool. It's often the clarity of the brief that we as humans struggle to create. This is exactly where AI literacy in schools should sit up and pay attention. Task imagination isn't a technical skill you learn by memorizing tool features. No, it's the ability to define a problem precisely to set clear quality criteria to anticipate failure modes and then to evaluate the output critically. Which is to say it's about thinking, the doing can increasingly be outsourced, the thinking that cannot. If anything, Fable V raises the premium on exactly the capabilities good teaching has always tried to build clarity of thought, sound judgment and the imagination to ask for something that might not even exist yet. It's about teaching students not to outsmart machines, but to outthink them, design and learning that cannot be faked because it demands depth, care and imagination. Now here's where the guardrails caught the science classroom, and it makes for a fascinating teachable moment. The launch wasn't all just wonder and impressive code migration, anthropic shipped fable five, with quite stringent safety classifiers. These are designed to detect requests that touch on sensitive areas like cybersecurity or anything related to biology and chemistry, or even attempts to copy the model itself. If a request falls into one of these buckets, the conversation is automatically rerouted to Anthropic's previous flagship model, Opus 4.8, and the user gets a notification. Anthropic says that over 95% of sessions don't involve any fallback at all, and they argue that falling back to Opus is a much better experience than a flat out refusal from Fable. In practice, though, the first few days were a bit bumpy. Users publicly shared examples of quite basic biology questions, including one about something as fundamental as mitochondria triggering this switch. The register, a technology news publication, even reported that one Inch Gates Foundation disease model and scientist found the classifier firing on essentially every session. Anthropic actually acknowledged pretty quickly that they'd made the safeguards too stringent and committed to reducing false positives, especially for biological research. They also said they'd make a previously invisible safeguard around frontier AI research visibly fall back to Opus as well. It would be very easy to just mock this, wouldn't it? To say, oh look, AI tripping over itself again. But I'd suggest that educators, especially those interested in teaching with AI, should read this very differently. Anthropic stated plainly that they released Fable with deliberately conservative, overly broad safeguards rather than delaying the launch. Why? Because the very same model that can accelerate legitimate scientific discovery can also unfortunately lower the barrier to misuse. That, my friends, is the dual use dilemma in action. And it's a crucial concept that every school teacher in AI literacy should be discussing with students. It's now visible in the very products they use every day. Imagine a biology teacher whose class watches a simple question about DNA, transcription, get rerouted. That's a live case study in AI governance sitting right there in the room. The teachable question isn't why is this annoying? It's much deeper. Why does this trade-off exist? Who decides where that line sits? And what would you do differently if you were in charge of setting these safeguards? It forces students to engage with ethics, judgment, and critical thinking, which are precisely the human domains AI cannot replicate. If you find these conversations helpful, please consider following this podcast. It helps us reach more educators who are grappling with how to effectively integrate AI into their practice. There's also a really practical note here for science departments. That fallback mechanism means that students and teachers, when asking life sciences queries, might actually be getting Opus 4.8 rather than Claude Fable 5. Now Opus is still a highly capable model, absolutely, but it's a timely reminder that the latest AI isn't a single uniform experience. The landscape is dynamic and understanding its nuances is part of that crucial AI literacy. Now, for school leaders, for anyone responsible for AI procurement and education, there are two further details that really deserve your careful attention. The fine print you need to read. First, let's talk pricing. Claude Fable V is included in paid Clawed Plans only until June 22nd. After that, access moves to usage-based pricing. And the API rates for Fable V are set at $10 per million input tokens and $50 per million output tokens, which is double the rate of Opus 4.8. What this means very clearly is that the era of flat rate access to frontier models appears to be ending. So schools that are planning their AI budgets need to start modelling usage-based costs rather than simply assuming that today's subscription terms are going to persist. This requires a shift in thinking for financial planning, tying into the start with why not how principle, understanding the actual use case and value, then budgeting appropriately. Second, and this is incredibly important for anyone handling student data, there's the question of data retention. Anthropic's documentation states that prompts and outputs for mythos class models are retained for 30 days for trust and safety purposes on every platform where they're offered. For organizations like schools that are handling sensitive student information, that 30-day retention window and the human review that it enables for trust and safety absolutely needs to be checked against your existing safeguarding policies, your data protection obligations, and any agreements you have in place. This isn't a reason to avoid the technology, not at all. It wouldn't is a reason though to read the documentation very, very carefully before the enthusiasm takes over. It's about being transparent about AI use and maintaining human accountability. So the question worth sitting with, the one that Fable V truly puts on the table, is profound. Every previous generation of AI models invited the same classroom question can students still do the work themselves? But Claude Fable V invites a better, more challenging one. Can students direct work they could never do themselves and then judge whether it's any good? That second question has always been the harder one to teach, hasn't it? It demands higher order thinking, judgment and critical evaluation, not just recall or execution. But it's also increasingly the one the world is going to pay for. The models will keep getting more capable. There's no doubt about that. But the institutions, the schools, the educational systems that thrive in this new landscape will be the ones that are developing people whose imagination, whose judgment, and whose standards grow just as fast as the AI itself. Schools have an incredible head start on that work if we choose to really see it and lean into it. The future of learning isn't just about what students can do, but what they can direct and judge. That's all for today. Thanks for listening.