Anthropic just put its most powerful public model in your hands, and then locked the dangerous twin in a vault next door.
On June 9, 2026, Anthropic launched Claude Fable 5, calling it the most capable model it has ever made generally available. Same day, it released Claude Mythos 5 to a tiny group of cyber defenders and infrastructure partners. Here’s the kicker: they are the same underlying model. The only thing separating them is the cage built around one of them.
Let’s break down what actually shipped, how good it is, and what it means for you whether you write code, run a business, or just open Claude to get work done.
What Fable 5 actually is
Anthropic introduced a new model tier called Mythos-class, which sits above its Opus class in raw capability. The first one, Claude Mythos Preview, launched back in April through a restricted program called Project Glasswing, available only to a handful of cybersecurity and critical-infrastructure providers. It was held back because it was simply too good at finding and exploiting software vulnerabilities.
Fable 5 is the version Anthropic decided was safe enough for everyone. Mythos 5 is the unrestricted twin, still locked behind trusted access. The name itself tells the story: Fable comes from the Latin fabula, meaning “that which is told,” which mirrors the Greek mythos. Same root, same model. What changes is the guardrails.
According to Anthropic’s own announcement, Fable 5 is state-of-the-art on nearly every benchmark it was tested on, with standout results in software engineering, knowledge work, vision, scientific research, and long-context tasks. The pattern worth remembering: the longer and more complex the task, the bigger Fable’s lead over older models.
The headline numbers that make this real
Benchmarks are easy to glaze over, so here are the concrete results that early testers reported:
- A two-month job done in a day. Payments company Stripe ran Fable 5 on a 50-million-line Ruby codebase. It performed a codebase-wide migration in a single day that would have taken a full engineering team over two months by hand.
- Drug design sped up roughly tenfold. Using the unrestricted Mythos 5, Anthropic’s internal protein-design experts accelerated parts of the drug-design process by around ten times, with the model matching or beating skilled human operators on certain protein targets.
- It beat a video game on vision alone. Older Claude models needed elaborate helper tools to play Pokémon FireRed. Fable 5 finished the game using nothing but raw screenshots.
- Memory that compounds. Given persistent file-based notes while playing the card game Slay the Spire, Fable improved its own performance three times more than Opus 4.8 did with the same setup.
The early customer quotes line up with the math. Cursor called it state-of-the-art on its internal coding benchmark and said it opened up long-horizon problems that were previously out of reach. A legal team reported its redlines matched or beat their current model every single time in blind review.
The catch: a model that sometimes hands you off
Here’s the thing that makes this launch different from any model release before it. Fable 5 ships with a set of safety classifiers, separate AI systems that watch for high-risk requests. When one fires, your request doesn’t get refused. Instead, it quietly gets answered by Anthropic’s next-best model, Claude Opus 4.8, and you’re told when that happens.
The classifiers cover three areas:
- Cybersecurity. Mythos-class models are frighteningly good at discovering and exploiting software flaws, so anything that smells like offensive hacking gets routed away.
- Biology and chemistry. Because the model can now do real scientific work, most bio and chem queries fall back to Opus 4.8 for now, to block any uplift toward bioweapons research.
- Distillation. Requests that look like attempts to copy Fable’s abilities to train competing models also get redirected.
Anthropic openly admits it tuned these safeguards too cautiously on purpose, so some harmless requests will get caught. The trade-off it made was clear: ship safely and fast, then loosen the filters later. The reassuring number is that early data shows fewer than 5% of sessions trigger a fallback. For the other 95% plus, you’re getting the full Mythos-class model with no compromise.
What it costs and how to get it
Both models are priced at $10 per million input tokens and $50 per million output tokens, which Anthropic says is less than half the price of the earlier Mythos Preview.
How you access Fable 5 depends on how you use Claude:
- Developers: Fable 5 is live right now via the Claude API using the model string
claude-fable-5. It supports up to 128k output tokens per request and is also available on AWS Bedrock and inside GitHub Copilot. - Subscribers: The rollout is staged. From launch through June 22, Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans. On June 23, it gets removed from those plans, and using it after that point will require usage credits until capacity catches up. Anthropic says it intends to restore Fable as a standard plan feature as soon as it can.
So if you’re on a paid plan, the practical move is simple: test it heavily before June 23 while it’s free.
The privacy change you should know about
One quieter detail matters for businesses. For Mythos-class models, Anthropic now requires 30-day data retention on all traffic, even for enterprises that previously had zero-retention agreements. The data is used to defend against novel attacks and reduce false positives, not to train new models, and it’s deleted after 30 days in nearly all cases. If your organization has strict data policies, factor this in before you route sensitive work through Fable 5.
Step-by-step: try Fable 5 today
- Pick your entry point. If you just want to chat, open Claude on a Pro, Max, or Team plan and select Fable 5 from the model picker. If you build software, grab the API string
claude-fable-5. - Throw a hard, long task at it. Fable’s edge grows with complexity, so don’t waste it on one-line questions. Give it a multi-step refactor, a long research synthesis, or a full document review.
- Watch for the fallback notice. If your request touches cybersecurity, biology, chemistry, or model copying, you’ll see a note that Opus 4.8 answered instead. That’s expected, not a bug.
- Use its memory. On long-running tasks, let it keep notes and reference them. That’s where it pulls away from older models.
- Move fast if you’re on a subscription. Run your real workloads before June 23 while access is free and unmetered.
Honest limitations
This is not a flawless release, and Anthropic doesn’t pretend it is.
- You won’t always get a Fable answer. Entire categories of legitimate work, especially in biology and chemistry, currently route to a weaker model.
- Capacity is uncertain. The June 23 cutoff for free subscription access exists precisely because Anthropic isn’t sure it can meet demand.
- Mandatory retention. The 30-day data rule is non-negotiable on this tier.
- It launched amid a warning. Anthropic released this days after publicly cautioning that frontier AI is advancing dangerously fast. The cage is the reason a public release was possible at all.
What this really means
Strip away the names and the story is striking. Anthropic built a model so capable that its unrestricted version was considered too dangerous to release openly. The thing that made a broad launch possible wasn’t a weaker model. It was a stronger cage. For most people, most of the time, that cage is invisible, and what’s left is the best assistant Anthropic has ever shipped to the public.
The future of AI might not be about who builds the smartest model. It might be about who builds the best lock.
Just Being Resourceful! 
Leave a comment