Anthropic announced Claude Fable 5 today. At first glance it is another new model in a long line of them. But this time something underneath breaks the normal rhythm.
Fable 5 is not a successor to Opus. It is the first time Anthropic has made a Mythos-class model broadly available, a tier that sits above the Opus class. In the announcement they call it “a Mythos-class model that we’ve made safe for general use.” Its capabilities exceed anything they have ever released to the public.
What was announced
The essentials.
- Two models, one core. Fable 5 is for everyone. Mythos 5 is the same model with the brakes off, for a small group of vetted cyberdefenders through Project Glasswing. Anthropic calls it the strongest cybersecurity capability of any model in the world.
- A built-in fallback. Fable 5 does not answer questions on cybersecurity, biology, chemistry, or distillation itself. Those quietly route to Claude Opus 4.8. It happens in under five percent of sessions, so more than 95 percent runs entirely on Fable. No other lab has shipped a designated fallback model as a launch feature.
- State of the art, broadly. Software engineering, knowledge work, vision, scientific research. On SWE-Bench Pro it scores 80 percent, against 59 percent for GPT-5.5. Its lead grows the longer and more complex the task.
- Long, autonomous work. The model can run autonomously for days in a harness like Claude Code. Stripe reports a 50-million-line Ruby migration done in a single day, work that previously took a team two months.
- Availability and price. Everywhere today: the Claude app, the API, AWS Bedrock, Vertex AI, and Microsoft Foundry. Ten dollars per million input tokens, fifty per million output. Included in Pro, Max, Team, and Enterprise through June 22, then on credits.
Why it feels different
The most interesting part does not come from the benchmarks, but from how the work changes. The Claude Code team put it sharply: “We used to verify that Claude did the work right. Now we verify that it’s doing the right work.”
In the accompanying video, someone from that team spells it out. He used to break every task into small chunks, check the output, and catch where the model stopped too early. With Fable 5 the model does that itself: it runs for hours, tests its own work, and by his own account sometimes writes better code than he does. His role becomes more about direction and setup than supervision. He names three things he now does differently:
- Claude as a thought partner. Bring it in while you are still thinking, not only at execution. He has the model interview him about a spec first, or asks for a few directions with quick HTML mockups before anything gets built. And he gives context, not bare constraints. Not “keep it simple,” but “this is an experiment, there’s a real chance we delete it in a month, so don’t build anything that would be painful to throw away.”
- Goals plus a way to verify them. He gives the model a goal and lets it check its own work, with a report of what was built and where it diverged from the plan.
- Be more ambitious. Ask for things you have never tried before. He edited the video itself with Fable.
That is a shift I recognize. When a model runs reliably for days, your role stops being about the individual steps. It becomes about the brief: am I asking the right question, am I describing the right goal. The work moves from doing to directing.
The reaction came from outside too. Aaron Levie, founder of Box, wrote: “If you thought AI progress was slowing down, here’s the immediate answer to that. Huge jump in capability across the board.” He expects it to push knowledge-work agents forward in almost every category. That covers most of what I run into in my sessions and my own work.
What I make of it
Two things stand out. The first is the fallback. Anthropic shipped a model so strong they built a second model as a safety net, so you do not always know which one answered you. That is a new pattern, and it says something about how seriously these capabilities are taken.
The second is that shift from doing to directing. That is where the real lesson sits for me. The tool gets so good that the value moves to the person who frames the right brief. That is exactly the work a release does not take over.
And I am going to put it to the test myself. Over the coming days I will run it through all kinds of situations: pulling out all my websites and apps and walking through them, maybe having an app rebuilt in a single shot, going over all the security, and watching the news around it closely. An announcement says something, a week of real work says more. What I find, you will read here later.
This piece also appeared in Dutch: Claude Fable 5: het sterkste model dat je nu mag gebruiken.