Anthropic did something today that it spent two months saying it might not do.
Back in April, they released a tier of models above Opus called “Mythos.” Mythos was so good at finding software vulnerabilities that Anthropic basically locked it in a vault, handed keys to a tiny group of cyber defenders through something called Project Glasswing, and told everyone else “not yet.” The reasoning was blunt: a model this capable, in the wrong hands, could do real damage.
Today they cracked the vault open. Sort of.
The new model is called Claude Fable 5, and Anthropic says its capabilities exceed any model they’ve ever made generally available. State of the art on nearly every benchmark they tested. And here’s the line that matters most: the longer and more complex the task, the bigger Fable’s lead. This is not a model built to win a chatbot popularity contest. It’s built to grind on a hard problem for hours, or days, without you babysitting it.
So why did I open by saying they tied a hand behind its back?
Because Fable 5 is the same underlying model as a second thing they released today, Mythos 5, which they are NOT giving to the public. The only difference between the two is safeguards. Fable is Mythos wearing a seatbelt. When you ask Fable something in a high-risk lane, cybersecurity, biology, chemistry, or attempts to copy the model itself, it doesn’t answer. It quietly hands your request to Opus 4.8 instead and tells you it did. Anthropic tuned those guardrails on the cautious side, so they admit some harmless questions are going to get caught in the net too.
Now, before you assume that makes Fable a watered-down disappointment, let me give you the realistic version. Anthropic’s own data says more than 95% of Fable sessions never hit that fallback at all. For almost everything most of us do, you get the full Mythos-class engine. The seatbelt only tightens in the lanes most people never drive in.
Let me show you what “full engine” actually means, because the numbers are genuinely a little ridiculous.
Stripe ran Fable on a 50-million-line Ruby codebase and reported that the model did a codebase-wide migration in a day. Their estimate for doing it by hand? A full team, over two months. Compress that down and sit with it for a second. Months of senior engineering work, finished while you get a coffee.
On vision, it’s the new state of the art. It can rebuild a web app’s source code from nothing but a screenshot. And the detail that made me laugh out loud: older Claude models needed an elaborate set of helper tools to play Pokémon FireRed, like a kid who needs the strategy guide open on his lap. Fable 5 beat the entire game using only the raw screen. No map, no navigation aids, no cheat sheet. It just looked at the screen and played, which frankly is more than I could do as an adult who still gets lost in Viridian Forest.
On memory, when researchers gave it a file-based notebook to track its own progress through a long game, its performance improved three times more than Opus 4.8 got from the same trick. That is the part that should make every founder and analyst lean in. Not the game. The “stays focused across millions of tokens and learns from its own notes” part. That is what turns a clever chatbot into something you can actually hand a project to.
So that’s the capability. Here’s the ground truth, the stuff nobody’s putting in the headline.
Thing one: for everyday users, Fable might feel a lot like Opus 4.8 more often than you’d expect. Because of how conservatively the safeguards are tuned, you’ll occasionally ask a perfectly innocent question and get bounced to the older model. Anthropic is upfront that this will be annoying and that they’re working to reduce the false alarms. Translation: the most powerful public model ever sometimes politely declines to be the most powerful public model ever. That’s the honest tradeoff for shipping it safely and fast.
Thing two: the “free” part has an expiration date. Read this carefully, because it’s the trap most posts will skip. On the Claude API and consumption-based Enterprise plans, Fable is fully available now, billed at usage. On subscription plans (Pro, Max, Team, seat-based Enterprise) it’s included at no extra cost from today through June 22. On June 23, it gets pulled from those plans, and using it after that needs usage credits. Anthropic says they aim to fold it back into subscriptions once they have the capacity, but they’re not promising a date. So if you’re on a Claude subscription and you want to actually feel this thing, your clean window is roughly two weeks. After that, the meter runs.
Thing three: the price tells you who this is really for. Fable runs $10 per million input tokens and $50 per million output. That’s less than half what Mythos Preview cost, which is a real price drop. But it’s still firmly in “serious work” territory, not “ask it to write a birthday card” territory. This is a model for the hardest, longest jobs you’ve got, the kind where you’d happily pay to compress two months into two days. For everything else, Opus is faster and cheaper, and Anthropic basically says so.
Thing four, and this is the one I keep circling back to. There’s a real data tradeoff. Anthropic now requires 30-day data retention on all Mythos-class traffic, including business customers, so they can catch novel attacks and jailbreaks. They say they won’t train on it and they delete it after 30 days. If you live in a world where data handling matters, and a lot of us building real products do, that’s not a footnote. That’s a line item for your compliance conversation.
Here’s where I zoom out, because this isn’t really a story about one model.
For two years the AI release pattern was simple: each new model is more capable AND more open than the last. Fable breaks that pattern on purpose. Anthropic is now shipping capability and restraint in the same box, with the restraint visible and admitted, not hidden. They literally gave the locked-down public version and the unlocked partner version two different names just to mark the line. The thing separating “anyone can use this” from “only vetted cyber defenders can use this” is no longer the model’s intelligence. It’s a layer of safety classifiers sitting on top of identical brains.
That’s a quiet but enormous shift in how this industry plans to scale. The frontier isn’t only getting smarter. It’s starting to come with a governor on the engine, and the companies building it are betting we’ll accept that tradeoff to get the power at all.
So what do you actually do with this? Honest answer:
If you’ve got a genuinely hard, long-running technical or analytical job, the next two weeks are a free test drive. Use them.
If you mostly do normal day-to-day work, Opus 4.8 is still your daily driver and you’re not missing much.
And if you’re building products on top of this, read the data retention policy before you read the benchmark chart. The benchmarks are exciting. The policy is what your lawyer cares about.
I’ll be honest about my own bias here: I came up in intelligence work, and the whole job was separating signal from noise under people who really, really wanted you to be impressed. A launch like this trips every one of those instincts. The capability is the signal. The two-week free window and the “it’s the most powerful model except when it isn’t” framing is the noise you have to read past.
Most powerful public model ever, on a free trial that ends in two weeks, that occasionally refuses to be itself. That’s the real headline. Everything else is marketing.
Sources below:
Anthropic — Claude Fable 5 and Claude Mythos 5 (official announcement): https://www.anthropic.com/news/claude-fable-5-mythos-5
Anthropic — Claude Fable product page: https://www.anthropic.com/claude/fable
AWS — Claude Fable 5 now available on Amazon Bedrock: https://aws.amazon.com/blogs/aws/anthropic-claude-fable-5-on-aws-mythos-class-capabilities-with-built-in-safeguards-now-available/
About Amazon — Claude Fable 5 from Anthropic now available on Amazon Bedrock: https://www.aboutamazon.com/news/aws/claude-fable-5-anthropic-available-amazon-bedrock
GitHub Changelog — Claude Fable 5 generally available for GitHub Copilot: https://github.blog/changelog/2026-06-09-claude-fable-5-is-generally-available-for-github-copilot/
CNBC — Anthropic releases Mythos-like AI model to the public, Claude Fable 5: https://www.cnbc.com/2026/06/09/anthropic-mythos-claude-fable-5.html
TechCrunch — Anthropic’s Claude Fable 5 is a version of Mythos the public can access today: https://techcrunch.com/2026/06/09/anthropic-released-claude-fable-5-its-most-powerful-model-publicly-days-after-warning-ai-is-getting-too-dangerous/
VentureBeat — Anthropic brings Mythos to the masses with Claude Fable 5: https://venturebeat.com/technology/anthropic-brings-mythos-to-the-masses-with-claude-fable-5-its-most-powerful-generally-available-model-ever
Anthropic — Project Glasswing: https://www.anthropic.com/glasswing
