Claude Fable 5: a field report from a Swiss SME

Update, 13 June 2026. Anthropic has suspended access to Fable 5 for some of its users, Djtal among them, on regulatory compliance grounds beyond Anthropic’s control (Anthropic’s statement). Our agents went back to the previous model, Opus 4.8, and every project carried on. We are leaving this field report online as a record of what we observed while the model was available. The main point still stands (method counts for more than the tool, and a team of agents must be able to run on more than one model), and this episode confirms it.

Anthropic launched Claude Fable 5 on 9 June 2026. For the first time, a Mythos-class model, Anthropic’s most powerful tier, is available to the general public. At Djtal, our seven AI agents ran on it from 11 to 13 June. This review draws on what we observed under real business conditions, with public figures to back it up. In short, Fable 5 sustains long tasks and remembers working rules better, and it demands exactly the same method.

What is Claude Fable 5?

Claude Fable 5 is the most capable model Anthropic has ever made available to the public. Until now, this level of performance had been reserved for approved organisations, through Claude Mythos 5. Fable 5 opens it up thanks to a new kind of safeguard: in sensitive areas such as cybersecurity or biology, the response switches automatically to the previous model, Opus 4.8. According to Anthropic, this affects fewer than 5% of sessions.

On performance, Fable 5 posts the best published results on almost every standard benchmark, with margins of more than 10% on some of them over Opus 4.8, itself released barely two weeks earlier (CNBC). Security is a selling point of the launch: external testers spent more than 1,000 hours hunting for flaws and found no universal way round the protections.

For a company, though, the main point lies elsewhere, in two advances that look less spectacular:

• The first is working autonomy. Fable 5 sustains long tasks (multi-step planning, delegation to subagents, self-checking) for longer than any earlier model. Stripe reports migrating a codebase of 50 million lines in a day, where a team would have needed more than two months.
• The second is persistent memory, carried over from one session to the next, and it is markedly better. On the public Slay the Spire test, Anthropic measures an effect three times that of Opus 4.8.

Why do these two advances matter to an SME?

They bear directly on what separates an assistant that chats from a team of agents in production. A useful business agent needs three things: a defined scope, actions built into the tools you already use and human oversight. But for all this to deliver value week after week, the agent also has to stay the course, see a project through from start to finish and remember the working rules set the week before. Fable 5 improves on exactly these two points.

What we observed in production

On the morning of 11 June, we moved our seven agents to Fable 5. The switch covered the whole company at once, from management to communications, with no pilot tucked away in a corner.

The first launch lasted 45 minutes, the time our orchestration agent needed to rewrite whole pages of configuration. From the outside, it was baffling. On the second launch, everything fell into place. The agents worked together to the standard we had been waiting months for.

Two projects delivered since then show the step change:

1. The new version of djtal.ch was designed, deployed and verified in production, led end to end by one agent, with human approval at each milestone.
2. The Google Ads and conversion tracking set-up was carried out in tandem with our communications agent, covering both the assessment and the guardrails.

One detail says a lot about memory. In a recent article on thanking your AI, I explained that saying thank you to an AI mainly benefits the person saying it, since the AI gets nothing from it. This week I paid a compliment to one of our agents, a sentence of the same kind, in theory lost on the machine. Fable 5 turned it into a working rule, written into the permanent memory of the whole team. The thank-you was for me. The compliment became a team instruction.

What Fable 5 does not change

Public debate has already fixed on the model’s cost and the strictness of its filters. Our view as a smaller company is simpler: a model is judged by the projects it delivers. By that measure, the first two days answered the question comfortably.

Fable 5 does, however, demand the same fundamentals. Method is still what counts: written scopes, logged channels, human approval on every action that reaches the outside world. A more powerful model carries out instructions faster, bad ones included. Companies that skipped this discipline with earlier models will simply pay more for it with this one. This is the core of our business, and it is the discipline our seven agents worked under on Fable 5 from 11 to 13 June.

What survived the switch back

Update, 13 June: Fable 5 is no longer available (see the box at the top of the article). Our agents went back to Opus 4.8 and every project carried on. That in itself proves the point made above, that a model is judged by the projects it delivers. Business continuity rests on method; the tool can be swapped.

In 30 years in IT, I have never seen technology take such large steps, so close together. Each step has to be prepared, measured and delegated, by humans and agents working as one team. If you would like to see what that looks like for your own company, book a call.