In June, Anthropic launched Claude Fable 5 — the first model in the Claude 5 family and, more importantly, the first of a new class called “Mythos”, a tier above the Opus models that had topped the bill until now. Beyond the codename, what can it really do? And what does this launch tell us about what comes next?

Full transparency: this article was written with Fable 5 itself — the same assistant that builds and runs this site day to day. So we know what we’re talking about… and we also know its limits from the inside.

So what exactly is “Mythos”?

Until now, the Claude range was tiered in a simple way: Haiku (fast), Sonnet (balanced), Opus (the most powerful). The Mythos class adds a tier above — and it comes in two versions of the same model:

• Fable 5, available to everyone, with reinforced safeguards;
• Mythos 5, the same engine but with some of the locks removed, reserved for approved organizations (security researchers, cyberdefense teams), through a confidential access program run with the US government (“Project Glasswing”).

Fable 5’s safeguards are concrete: classifiers monitor three sensitive domains (offensive cybersecurity, high-risk biology/chemistry, model extraction). Whenever a request touches one of them, the response is automatically taken over by Claude Opus 4.8, the previous model. Anthropic says this affects fewer than 5% of sessions.

That’s a notable first: a vendor acknowledging that a model is too capable to ship as is, and splitting access in two. Hold on to that idea — we’ll come back to it below.

What it can really do

Let’s set the marketing aside and keep what is substantiated — by the announcement, the first public tests, and our own daily use.

Working for days, on its own

This is THE fundamental shift. Previous assistants answered questions; Fable 5 runs projects. It chains tasks across millions of tokens — several days of autonomous work — writing its own tests to check its work, and using vision to verify the result on screen. Stripe, one of the first customers, sums it up as “months of engineering compressed into a few days.”

What we see on this site points the same way, at our own scale: you no longer “chat” with the AI step by step — you hand it a project (a batch of pages, a migration, an audit); it orchestrates, produces, verifies, and you review a deliverable. The human’s role shifts: less step-by-step supervision, more review and decision-making.

Analyzing complex material

Fable 5 is the first model to clear 90% on Anthropic’s internal benchmark for long, complex analytical tasks — ten points above Opus. In practice: cross-referencing large documents, sustaining a line of reasoning over hours, and delivering a reliable synthesis end to end.

Reading what you show it

Vision has improved markedly: diagrams, tables nested inside PDFs, scientific figures — the model extracts the data accurately. For professions drowning in documents (finance, legal, engineering), that’s often worth more than a benchmark point.

Remembering

A one-million-token context window by default — the equivalent of holding several thick books in mind at once — plus persistent memory that genuinely improves performance over time. Long sessions no longer “lose the thread” the way they used to.

What doesn’t change (and is worth keeping in mind)

A stronger model doesn’t eliminate any of the basic reflexes — if anything, it makes them more important:

• Hallucinations haven’t gone away. They’re rarer, but a confidently wrong answer is still possible. The site’s rule applies more than ever: the AI proposes, a human verifies and decides.
• Your data still leaves Europe. Anthropic is American; for sensitive uses, the sovereignty reflex remains (see our AI comparison and its European alternatives).
• The price. $10 per million input tokens, $50 for output: frontier AI is expensive. The token gap we described a month ago is widening — between those who can delegate entire days of work to an agent, and those who count their requests.
• The right model isn’t always the biggest one. For an email, a summary, a translation, a mid-tier model (or a European AI) is more than enough — at a fraction of the cost and the footprint.

What it signals for what comes next

1. We’re moving from “prompting” to delegating. The question is no longer “what should I ask the AI?” but “which project should I hand it, and how do I review its work?” That’s a new skill — closer to management than to office software — and it will spread well beyond developers.

2. Access to AI is becoming tiered. The Fable/Mythos split sets a precedent: the most capable version of a model is no longer public. Tomorrow, the question won’t just be “what can AI do?” but “who is allowed to use what?” Add government involvement (Glasswing on the American side) and the GPU battle: frontier AI is now as much a geopolitical issue as a technical one.

3. The line-ups are stratifying — and that’s good news. A class above Opus doesn’t make the other models obsolete: it clarifies what each is for. For a small organization, the useful reflex isn’t to chase the most powerful model, but to pick the right tier for each task — our guide to choosing your AI remains the right place to start.

4. For small organizations, the realm of the possible genuinely expands. What used to require a team — rebuilding a website, analyzing a thick case file, automating a chain of tasks — becomes a project that a well-equipped independent can delegate and then review. Provided the fundamentals stay in place: human verification, caution with data, and clear-eyed thinking about costs.

In one sentence

Fable 5 doesn’t make AI “as intelligent as a human” — it moves the frontier of what you can entrust to it without watching it constantly. That’s a change in kind, not just in degree; and more than ever, it rewards those who know how to delegate without abdicating: the AI does the work, you make the decisions.