The Complete Guide to Claude Fable 5

On 9 June 2026, Anthropic released Claude Fable 5 — the first Mythos-class model available to the public. In plain terms: Anthropic builds research-grade frontier models internally (Mythos), and Fable 5 is the first one they’ve made safe enough to hand to everyone. It sits a full tier above the Opus models that were, until last week, the best you could get.

Three things matter for a business owner:

It holds your whole business in its head. Fable 5 can take in about a million tokens of context — roughly 2,000 pages of documents — in a single conversation. Your full CRM export, a year of campaign data, your supplier contracts and your brand guide, all at once, no summarising.

It finishes long jobs on its own. The headline improvement isn’t trivia or chat — it’s what Anthropic calls long-horizon work. Anthropic’s own line: “the longer and more complex the task, the larger Fable 5’s lead.” Give it a project, not a question, and the gap over every other model widens.

It’s briefly cheap, then premium. Through 22 June it’s included in Pro, Max, Team and Enterprise plans at no extra cost (burning your plan limits about twice as fast as Opus 4.8). From 23 June it leaves those plans and costs usage credits at US$10 per million input tokens and US$50 per million output — double Opus 4.8. Anthropic says it aims to bring it back into subscriptions when capacity allows, but there’s no date on that.

One honest footnote: on a small set of sensitive topics (security exploits, dangerous biology — under 5% of sessions), Fable 5 quietly hands your request to Opus 4.8 instead. For business use you will almost certainly never notice.

Benchmarks aren’t everything, but they’re the closest thing this industry has to a measuring tape. Every figure below was checked against its source on 10 June 2026, and we’ve labelled who did the measuring. A handful of viral numbers floating around social media didn’t survive verification, so you won’t find them here.

That last chart is the story. On everyday tasks Fable 5 is a solid step up. On the hardest, longest tasks, it’s more than double its own predecessor — which was already the best in the world.

Independent verification: the analysts at Artificial Analysis ran their own evaluations and ranked Fable 5 #1 of 152 models on their Intelligence Index — about five points clear of GPT-5.5.

Worth knowing: it’s smarter, not faster. Output speed is mid-pack (~60 tokens/second per Artificial Analysis), and it spends real time thinking before it answers. Fable 5 is for jobs where you care about the result, not the stopwatch.

You can actually play some of this. Mollick commissioned a complete game and an interactive NYC travel-time map from single prompts — click through and poke them.

Closer to home for business owners: marketer Simon Bradley tested five real-world business prompts — it built him an interactive business simulator in under three minutes and rewrote corporate copy into three distinct brand voices. His honest caveat, which I’ll echo: for routine work, the regular models are enough. Fable 5 is for the hard stuff.

And a ninth case study: this page. Spec’d, researched by parallel agents, fact-checked, written, and deployed the same day — with Fable 5 doing the heavy lifting.

This is the deep section. Everything above tells you what Fable 5 is; this tells you how to actually work with it. The single biggest mistake people will make with this model is using it like the old ones.

Stop chatting. Start commissioning.

Mollick’s “commissioning a small studio” line is the right mental model. With previous models, you had a conversation: ask, answer, refine, ask again. With Fable 5, you write a brief — the way you’d brief a contractor — and let it work. The model is specifically built to take a big, clearly-specified job and run with it. Drip-feeding it context one message at a time wastes exactly the thing you’re paying (or about to be paying) for.

Give it the whole job up front

A good Fable 5 brief has the same bones as a good contractor brief: the goal (what done looks like, concretely), the context (every relevant document — remember, it can hold 2,000 pages), the constraints (budget, brand rules, things it must not do), and the format you want back. Anthropic’s own guidance for this model class says the same thing: full task specification up front, in one well-specified go, beats twenty clarifying messages.

Then let it run

The benchmark story — double the score on the longest tasks — only pays off if you give it long tasks. Resist the urge to interrupt every two minutes. Testers ran it for hours unattended; Mollick’s big run went nine and a half hours, with the model organising its own sub-helpers along the way. Set the job going, make a coffee, review the result like a director reviewing a cut.

It decides how hard to think — you don’t

Older models had settings to fiddle. Fable 5 uses what Anthropic calls adaptive thinking: simple ask, quick answer; hard ask, it visibly slows down and reasons. There’s nothing to configure. Your lever isn’t a setting — it’s the quality of the brief.

Know when NOT to use it

Three cases. Routine work: for emails, summaries and everyday questions, the standard models are excellent and don’t burn your limits at double speed — Bradley’s “Opus is enough” caveat is correct. Speed-critical work: Fable 5 thinks before it talks; if you need instant back-and-forth, it’s the wrong tool. After June 22, small jobs: at $50 per million output tokens, you reach for it the way you’d call a senior specialist — for the work that justifies the rate.

The beginner mistakes, in one list

Treating it like Google (it’s a worker, not a search box). Drip-feeding context (give everything up front). Skipping review (it is brilliant and still wrong sometimes — you are the editor). Burning the window on small tasks (everyday models exist). And not writing down what worked — the brief that produced gold is an asset; save it, reuse it, refine it.