Claude Fable 5 Explained

Six jobs to give Fable 5 — copy-paste prompts

These are the jobs where the extra horsepower actually shows up — deep review work across your whole setup, the kind a smaller model skims. Paste each prompt into Claude Code (or Cowork) with Fable 5 selected, then go make a coffee. (Heads up: since 7 July Fable needs usage credits — these six are exactly the jobs worth spending them on. Or run them on Opus 4.8 first, free, and only upgrade the ones that fall short.)

1

Get the handoff docs

Fable is the best employee you only hire occasionally — so ask it to write its own handoff docs. This one prompt gets you a ranked plan of the 10 skills worth building while you have its brain.

You're my best employee and you've just handed in your notice — before you go, I want your handoff docs. Read everything that shows how I actually work: my CLAUDE.md or agents.md files, my memory, my project docs, and our recent chats. Don't sample — read enough to know what I do every week. Then give me the top 10 skills you'd build for me. Skills are files of repeatable instructions — SOPs for my AI. For each one tell me: 1. The name and what it does, in one line. 2. What I currently do by hand that it replaces. 3. Roughly how much time it saves me per week. This is the plan only — don't build anything yet. Rank the list by time saved and flag the two or three you'd build first. When I've picked, build the ones I choose one at a time and show me how to trigger each. Bonus, once the builds are done: update my CLAUDE.md or agents.md with anything you learned about how I work, so the cheaper models plan the way you would.

2

Audit every skill you've built

Your skills were written on different days, in different moods, some by an older model. Fable reads the lot in one sitting and finds the quietly broken ones.

Audit every skill in my skills folder, one by one — don't sample, read them all. For each skill tell me: 1. What it does, in one line. 2. What's weak — vague instructions, missing edge cases, references to tools or files that no longer exist, steps that rely on me remembering something. 3. Whether it overlaps or conflicts with another skill. Then rank the full list from "working hard" to "quietly broken", pick the three that would save me the most time if improved, and rewrite those three in full. Show me a before/after for anything you change. Don't invent new skills — improve what exists. I have to maintain all of this myself, so simpler beats clever.

3

Run a security audit on your own setup

This is the job Fable is genuinely best-in-class at — security review is the exact capability Anthropic put the extra guardrails around. Point it at your own business.

You're doing a security review of my own setup — everything you can see in this workspace belongs to me and I'm asking you to check it. Look for: - API keys or passwords sitting in files they shouldn't be in (hardcoded in scripts, committed to git, pasted into notes) - Webhook URLs or form endpoints that anyone on the internet could hit - Anything I've made public that shouldn't be — test pages, staging links, open Google Sheets or Docs - Permissions and connections that are wider than they need to be - Automations that would do something expensive or embarrassing if a stranger triggered them For every finding, give me: what the risk is in plain English, how likely it actually is, how bad it would be, and the exact fix. Sort into "fix today", "fix this week", and "nice to have". Don't change anything yourself — report first, I'll say go.

4

Review your whole Claude setup

Memory, instruction files, skills, connections, folder structure — the stuff that accumulates. Fable holds all of it in its head at once and tells you what to prune.

Review my entire Claude Code and Cowork setup like a consultant who charges by the improvement — my CLAUDE.md files, memory, skills, hooks, connected tools, and folder structure. Tell me: 1. What's working that I should lean on harder. 2. What's redundant or fighting itself — duplicate instructions, stale memory entries, skills that overlap or contradict each other. 3. What's missing that someone with my workflows would clearly benefit from. 4. The single highest-leverage change, if I only make one. Be specific — name the actual files. For every suggestion, give me the exact edit rather than general advice. Assume I maintain everything myself with no developer, so anything clever that I won't understand in six months is a downgrade, not an upgrade.

5

The money-leak audit

Fable's document upgrade means it can actually read bank statements, receipts and messy PDFs properly. Feed it a year of expenses and let it find the leaks.

Here are my bank and card statements plus subscription receipts for the last 12 months [attach the files or point it at the folder]. Go through all of them — don't sample. Find: 1. Subscriptions I'm paying for twice, or tools that overlap with each other. 2. Prices that have quietly crept up since I signed up. 3. Anything I clearly haven't used in three months or more — ask me if you can't tell. 4. Charges I probably wouldn't recognise if you read them out to me. Give me one table sorted by annual cost, with a keep / cancel / downgrade call for each line and the total I'd save per year if I took every recommendation. Plain English, no finance jargon.

6

Get the business out of your head

Fable can hold your whole workspace in its head at once — perfect for writing down the processes that currently exist only in your brain. This is the one that pays off every week after.

Read through my workspace — project folders, docs, and our chat history — and work out the five business processes I clearly run every week but have never written down. For each one, write a proper SOP: 1. The trigger — what kicks the process off. 2. The steps, in order, exactly as I actually do them (not the idealised version). 3. The tools and files involved, with their real names and locations. 4. What "done" looks like. Write them so a new VA — or a cheaper AI model — could follow them without asking me anything. Save each as its own file and give me the list. If you're guessing on any step, mark it with [CHECK] instead of inventing details.

Why these six?

They're all review jobs across a big messy surface — exactly where a stronger model earns its keep. A caption doesn't get better with a smarter model. An audit of forty files does. Run them once, keep the reports and the skills plan, and you've got a fix-it list that pays off for months.

What just happened

On 9 June, Anthropic released Claude Fable 5 — the public version of Mythos, a model it had kept locked away since April because of what it could do in cybersecurity.

1

April 2026 — Mythos unveiled, then restricted

Anthropic announced Mythos but limited who could use it, citing its advanced cybersecurity capabilities. Most of us never got to touch it.

2

April–June — quiet expansion

Access slowly widened to hundreds of vetted critical-infrastructure organisations across 15 countries. Still nothing for the public.

3

9 June 2026 — Fable 5 goes public

Anthropic released Fable 5: the same Mythos-class model, but with safety guardrails bolted on so the rest of us can use it.

The plain-English translation

Fable 5 is Mythos with a seatbelt. Same engine underneath — Anthropic has confirmed it's the identical model — just with filters that block the dangerous stuff. It beats every Claude model you've used before on nearly every benchmark, including more than 10% higher than Opus 4.8 on some tests.

If you want the source material: Anthropic's announcement has the full detail.

What it can actually do

These aren't marketing claims — they're verified results from named companies and third-party benchmarks. Here's what's genuinely new, translated out of tech-speak:

🏗️

Months of dev work in days

Stripe used it to complete a code migration — 50 million lines — in one day. Their team had budgeted two months. This is the headline capability.

🏃‍♀️

Works unattended for days

Previous models lost the plot on long jobs. Fable 5 can keep working on a complex task for days without a human nudging it along.

📄

Actually reads messy documents

Charts, diagrams and tables buried inside PDFs — the stuff older models skimmed past or misread — it now understands properly.

📊

Serious financial analysis

First model to break 90% on Hebbia's finance benchmark — senior-analyst-level reasoning across long, complex analytical tasks.

🧠

Remembers across long jobs

With file-based memory, its performance improved 3x more than Opus 4.8 on extended tasks. It holds context instead of forgetting what it did an hour ago.

👀

Sees like a person

It rebuilt working web apps from nothing but screenshots, and played a Pokémon game using vision alone. Show it a thing; it understands the thing.

What that means in normal business terms

If your work involves huge piles of documents, long-running analysis, or genuinely big builds, this model is a real step up. If your work is writing, planning, client comms and everyday automations — keep reading, because the next two sections matter more than this one.

What it can't (and won't) do

Remember, this is the model Anthropic considered too risky to release in April. The public version ships with guardrails — safety classifiers that watch what you ask and block certain territory entirely.

Offensive cybersecurity — hacking, exploits, anything attack-shaped

Dangerous biology and chemistry questions

Attempts to copy its capabilities to train rival models

Your marketing, ops, content, finance and admin work — untouched

When a guardrail triggers, Fable 5 doesn't just refuse — it quietly hands your question to Opus 4.8, which answers instead. Over 95% of sessions never trigger one at all. The filters are deliberately cautious though, so if you work in health, science or security-adjacent fields, expect the occasional false positive.

The catch nobody is talking about: 30-day data retention

Every Fable 5 conversation is kept by Anthropic for 30 days — mandatory, for everyone, even enterprises that previously had zero-retention agreements. It's not used to train models and access is logged, but if you've made strict privacy promises to clients about their data, factor this in before running client material through it.

The price tag nobody mentions

The launch posts are all about benchmarks. Here's the part that affects your wallet.

💸 Claude Fable 5

$10 USD per million input tokens
$50 USD per million output tokens
Double the price of Opus 4.8
Since 8 July: separate usage credits on top of your subscription

✅ Claude Opus 4.8

Half the price, via API
Included in your existing paid plan
Already excellent at business writing, analysis and automations
Still the fallback model Fable 5 itself leans on

Quick jargon check: a token is roughly three-quarters of a word — "input" is what you send, "output" is what Claude writes back.

How the rollout worked on paid plans

9 June – 7 July

Fable 5 was included at no extra cost on Pro, Max, Team and Enterprise plans. That free window has now closed.

Now

It needs separate usage credits — real extra money on top of your subscription.

Anthropic says it plans to fold Fable 5 back into standard plans "as quickly as we can" — no date given. Translation: if you wait, you'll likely get it included anyway.

Why you probably don't need it

Time for the honest bit. You use Claude for emails, captions, client documents, automations and planning. Fable 5 wins at 50-million-line code migrations, genomics research and days-long engineering runs. Notice the gap? The benchmarks it dominates are nowhere near most business owners' daily work.

The uncomfortable truth about benchmarks

A model that's 10% better on a frontier coding benchmark will not write a 10% better welcome email. Your everyday tasks stopped being limited by model intelligence a few releases ago — Opus 4.8 and Sonnet 4.6 already sit well past the "good enough" line for business writing, planning and automations. Past that line, extra capability is invisible.

You'd be paying double for a difference you genuinely couldn't pick in a blind test.

The three-question filter (use it for every model launch, ever)

Does my actual work hit the current model's ceiling? If Claude never feels "not smart enough" for your tasks, a smarter model fixes a problem you don't have.
Would I notice the difference blind? If you couldn't tell which model wrote the caption, don't pay extra for the fancier one.
Am I paying for capability or novelty? New-model excitement is real and fun. It's also not a business case.

Fable 5 isn't overhyped — it was just built mostly for engineers, researchers and analysts. That might occasionally be you. It isn't you every day.

When you WOULD reach for it

There are real moments in a service business where Fable 5 earns its price tag. They're the "big lift" jobs — the ones you currently break into pieces because the model loses the thread, or avoid entirely because they're too tedious.

1

Giant document analysis

A year of financials, a stack of supplier contracts, 40 client onboarding docs — anything where charts and tables live inside PDFs and accuracy actually matters.

2

Long unattended runs

A full audit of your website, content library or funnel in one continuous session — the kind of job where older models needed babysitting every twenty minutes.

3

The once-a-year big build

Migrating platforms, rebuilding a site, untangling years of accumulated systems mess. Rare, painful, and exactly what this model was built for.

The cheat sheet: use this model for this job

A

Everyday writing and quick tasks — captions, emails, brainstorms, summaries → Sonnet 4.6. Fast, cheap, more than smart enough.

B

Business thinking and building — strategy docs, automations, client deliverables, code via Claude Code → Opus 4.8. Your daily driver. Included in your plan.

C

The big lifts — massive document piles, days-long jobs, the once-a-year build → Fable 5. Worth the credits when the job is genuinely huge.

What to do now

The free window closed on 7 July — Fable 5 now needs usage credits on top of your subscription. That's not bad news, it just changes the play:

1

Don't buy credits for everyday work

Captions, emails and automations won't get better. Opus 4.8 is included in your plan and already sits past the "good enough" line.

2

Save it for the big lifts

When a genuinely huge job lands — a platform migration, a year of messy financials, a full-business audit — a small credit top-up spent on the prompts at the top of this page is money well spent.

3

The rematch test (before you spend)

Find the one task where Opus genuinely disappointed you and run it again on Fable 5, same prompt. If you can't pick the winner blind, keep your money.

The decision rule

If a job doesn't pass the three-question filter above, skip the credits with zero guilt — Anthropic has said it intends to fold Fable 5 back into standard plans anyway. Paying double for a difference you can't pick in a blind test isn't an investment, it's a souvenir.