Anthropic Fable 5, Explained Simply: The New Most-Capable AI Model

On 9 June 2026, Anthropic (the company behind Claude) released a new AI model called Fable 5. They call it their most capable model yet. If you run a business, your first question is the right one: what is it, and does it change anything for me? This is the plain-English answer, with pictures. Read the four-minute version first. If you want the full tour, there is a button at the end that opens it.

The one-sentence version

Fable 5 is a smarter, more careful version of the AI you may already know, built for the hardest jobs (long, multi-step work where being right matters more than being cheap), and it costs more to match.

The headline numbers

Three numbers tell most of the story. Do not worry about the jargon yet, the deep dive unpacks each one.

1Mtokens of memoryIt can read and hold a very large pile of documents at once.

128Ktokens of outputIt can write a long answer in one go without running out of room.

$10 / $50per million tokensThe cost to read ($10) and to write ($50) a million tokens. Premium tier.

Where it sits in the lineup

Anthropic does not sell one AI. It sells a small range, from light and fast to heavy and powerful. Think of it like engines. Fable 5 is the new top of the range, a tier they call "Mythos-class" that sits above the Opus models.

The Claude lineup, lightest to most capable. Fable 5 is the new top tier. More capable also means more expensive, so the right pick depends on the job.

What it means for your business

You do not buy a model. You buy what it lets you build. Here is the honest read for an owner.

It widens what is possible, not what is necessary. Jobs that were too hard or too unreliable for AI a year ago are now in reach: working through a whole messy process end to end, not just a single step.
It is a premium tool. At $10 to read and $50 to write per million tokens, you do not run your high-volume, simple tasks on it. You point it at the few jobs where a better answer is worth real money.
Most businesses still start cheaper. A lighter model handles the bulk of everyday work. Fable 5 earns its keep on the hard twenty percent. Choosing the right one for each job is the whole game.

That is the four-minute version. The deep dive below takes it apart piece by piece: what a "model" even is, what those numbers really mean, how it thinks, how it stays safe, what it has actually done, and a simple guide to which model to use when. No prior knowledge needed.

Read the full 20-minute deep diveEvery part of Fable 5, explained from scratch

1. First, what is an AI "model"?

A model is the engine inside an AI assistant. It is a very large piece of software that has read an enormous amount of text and learned the patterns in language well enough to answer questions, write, summarise, translate, and follow instructions. When you type into a chat assistant, a model is the thing on the other end working out the reply.

"Claude" is Anthropic's family of these models. Each new release is smarter and more reliable than the last. Fable 5 is the newest and strongest member of that family. You do not need to understand how it works inside, the same way you do not need to understand a car engine to drive. You need to know what it is good at, what it costs, and where to point it.

2. The family tree, in plain terms

Anthropic releases models in tiers. Lighter models are fast and cheap. Heavier models are slower and dearer, but they reason better and make fewer mistakes on hard problems. Fable 5 introduces a new top tier called "Mythos-class," which sits above the Opus tier that was previously the strongest.

Model	Tier	Best for	Who can use it
Haiku 4.5	Light	High-volume, simple tasks where speed and price win	Everyone
Sonnet 4.6	Balanced	Most everyday production work	Everyone
Opus 4.8	Heavy	Hard reasoning at a lower cost than the top tier	Everyone
Fable 5	Mythos-class	The hardest, highest-stakes, long-running work	Everyone
Mythos 5	Mythos-class	Frontier science (same engine as Fable 5)	Gated: research partners only

Fable 5 and Mythos 5 are siblings built on the same engine. The difference is access. Anyone can use Fable 5. Mythos 5 is reserved for Anthropic's Project Glasswing partners and a small group of biology researchers working on sensitive frontier problems. For a normal business, Fable 5 is the one that matters.

3. "State of the art": what has it actually done?

Anthropic describes Fable 5 as state of the art on nearly every benchmark they tested. Benchmarks are standardised exams for AI models. Anthropic did not publish a table of exact scores in their announcement, so I will not put invented percentages on this page. What they did share are concrete, real-world results. These are Anthropic's own reported numbers, worth knowing and worth taking as claims from the maker.

Area	What Anthropic reports
Software engineering	At Stripe, a 50-million-line code migration that was estimated at two months was completed in a single day.
Drug design	Sped up parts of the drug-design process by around ten times.
Vision	Best-in-class at reading scientific figures and rebuilding a working web app from a screenshot.
Coding (FrontierCode)	Highest score among frontier models on Cognition's evaluation.
Finance (Hebbia)	Highest score of any model on the Hebbia Finance Benchmark.

The pattern to notice: the wins are on long, hard, multi-step jobs. That is the kind of work Fable 5 is built for, and it is the reason it costs what it does.

4. The 1M context window, explained

"Context window" is the amount of text a model can hold in mind at one time. Picture it as the size of its desk. A small desk fits a page or two. A huge desk fits a stack of binders, all open, all readable at once. Fable 5 has a one-million-token desk.

A "token" is roughly three-quarters of a word. So one million tokens is in the region of 750,000 words. For comparison, a typical business novel is around 90,000 words. That means Fable 5 can read the equivalent of several long books in a single sitting and answer questions across all of them at once.

The context window is the model's working memory: everything it can read and reason over in one go. A larger window means it can hold your whole contract, codebase, or document set at once, rather than in forgetful chunks.

Why this matters for you: a big window means you can hand the model your whole contract, your entire product manual, or a year of support emails and ask it to reason across all of it, instead of feeding it small pieces and hoping it remembers. Fewer pieces means fewer dropped details.

5. 128K output: how much it can write back

The context window is what it can read. The output limit is what it can write in a single reply. Fable 5 can produce up to 128,000 tokens at once, roughly 96,000 words. In practice that means it will not cut off halfway through a long report, a full rewrite, or a big batch of generated content. You rarely need that much, but having the headroom means long jobs finish in one pass.

6. "Always-on thinking" and the effort dial

Older AI models answered instantly, like blurting out the first thing that comes to mind. Newer ones can "think" first: work through a problem privately before replying, the way a careful person pauses before answering a hard question. On Fable 5 this thinking is always on. You cannot switch it off, because it is part of what makes the answers good.

What you can control is how hard it thinks, using a setting called effort. More effort means deeper reasoning and a better answer on hard problems, at the cost of more time and more tokens. Less effort is faster and cheaper for simple work.

The effort dial trades speed and cost against depth of thinking. 'High' is the standard setting; 'xhigh' is the sweet spot for hard, agentic work; 'max' is for when being right matters more than the bill.

For a business, the takeaway is simple. The dial lets whoever builds your tool tune the trade-off per job: low effort for routine sorting, high or xhigh for the work that needs a careful answer. You pay for the thinking you actually use.

7. Token efficiency and a new way of counting

Two things are true at once here, and they sound contradictory until you separate them.

Fable 5 is more token-efficient at getting work done. It tends to reach a good answer with less wasted back-and-forth than older models. Smarter per token.
It counts text with a new "tokenizer." The same sentence is measured as roughly thirty percent more tokens than on the older models. So if you were budgeting based on old token counts, re-measure, do not assume the old maths carries over.

The same piece of text is counted as about 30% more tokens on Fable 5's new tokenizer. This is a change in how text is measured, not a sign the model is wasteful. Re-baseline any cost estimates rather than reusing old numbers.

8. Safety, built into the model

A more powerful model needs stronger guardrails, and Anthropic built them in rather than bolting them on. For a small set of genuinely sensitive subjects (cybersecurity, biology and chemistry, and a few others), Fable 5 does not answer directly. The request is quietly handed to a different model, Opus 4.8, or politely declined. Anthropic reports this happens in under five percent of sessions, so most users never notice it.

For most requests, Fable 5 answers directly. For a small set of sensitive topics, a safety layer routes the request to Opus 4.8 or declines, by design. This is a normal, expected branch, not a failure.

Anthropic also stresses how hard it resisted attempts to trick it. They report no universal jailbreaks found in over a thousand hours of outside testing, and zero harmful responses on a battery of cybersecurity tricks. For a business, that is the reassurance that the tool will not be easily talked into doing something it should not.

9. Long jobs: it can work for many minutes

Most AI replies come back in seconds. Fable 5 is different on hard tasks: it may genuinely work for many minutes, planning, trying, checking, and correcting itself, the way a person tackles a big job. Two features keep that under control.

Two tools keep long jobs sane. A task budget lets the model watch how much it has spent and pace itself. Compaction folds a long conversation into a summary so it can keep going past the memory limit.

You will not touch these settings yourself, your build team does. But it is worth knowing the model is designed to run long, careful jobs without losing its place or running away with your budget.

10. Getting access, and what it costs

Fable 5 is generally available, meaning anyone can use it. It is in the Claude apps and through the developer interface that powers custom tools. On the paid Claude plans (Pro, Max, Team, Enterprise) it was rolled in from 9 to 22 June 2026, and after 23 June heavier use draws on usage credits. Mythos 5, its frontier sibling, stays gated to research partners.

What	Fable 5
Cost to read text (input)	$10 per million tokens
Cost to write text (output)	$50 per million tokens
Versus the previous frontier model	Less than half the price of the earlier Mythos Preview
Data handling	30-day retention; your data is not used to train the model

To put the price in everyday terms: a million tokens of input is around 750,000 words. So reading the equivalent of several long books costs in the region of ten dollars, and a long written answer is priced higher because generating text is the expensive part. It is a premium tool, and it is meant to be used like one.

11. Which model should you actually use?

This is the question that matters, and the answer is almost never "always use the most powerful one." The skill is matching the model to the job. Here is a simple guide.

If the job is...	Reach for	Why
Simple and very high volume (sorting, tagging, short replies)	Haiku	Speed and price win; the task does not need deep reasoning
Everyday production work	Sonnet	The balanced default for most real tasks
Hard reasoning, but cost-sensitive	Opus 4.8	Strong reasoning without the top-tier price
The hardest, highest-stakes, long-running work	Fable 5	When a better answer is worth real money and being wrong is costly

A well-built system often uses more than one: a cheap model for the bulk, the heavy model only on the steps that need it. That is exactly the kind of decision we help businesses get right, so you are not paying top-tier prices for bottom-tier work.

12. The honest limits of this article

A note in the spirit of straight talk. Anthropic did not publish exact benchmark percentages with this release, so you will not find invented scores here. The real-world results above (the Stripe migration, the drug-design speedup, the finance and coding wins) are Anthropic's own reported figures, which is to say they are claims from the maker, credible but not independently audited on this page. When independent numbers land, they are the ones to weigh. Treating a maker's announcement as the maker's announcement, not as settled fact, is just good judgement.

The bottom line

Fable 5 is a real step up in what AI can reliably do on hard, long, high-stakes work, at a premium price to match. For most businesses it is not the everyday workhorse, it is the specialist you bring in for the jobs that justify it. The win is not "use the newest model." The win is knowing which model to use where, and building so each job runs on the cheapest one that does it well. That is the part worth getting right, and it is the part we are happy to help with.

Source: Anthropic's official announcement (9 June 2026), plus the published Claude API model reference. Pricing and access details are from Anthropic and may change.