← Back to recent questions

Fable is Anthropic's most capable model to date, now available to everyone for the first time. It sits above Opus in a new tier Anthropic calls Mythos-class, meaning it can do things no publicly available Claude model has done before. Whether it is worth using depends on the type of work you are doing, and before you switch there is a privacy change that is genuinely important to understand, especially if your work involves anything confidential.

Where does Fable sit, and why does a new tier exist?

Until three days ago, Opus was Anthropic's most capable public model. Above Opus there was a tier called Mythos, but Mythos was only available to a small group of vetted partners, mainly cybersecurity organisations working with governments. Fable is Anthropic's first step in bringing Mythos-level capability to the general public.

The reason a new tier exists at all is capability, not marketing. Mythos-class models can perform autonomous scientific reasoning, find and exploit software vulnerabilities at a meaningful level, and work through genuinely long and complex tasks without losing the thread. That combination of capability is powerful enough that Anthropic held it back until it could add safety classifiers, which are separate AI systems that monitor requests and redirect certain types of query. More on those below. Fable is essentially Mythos with those classifiers running in front of it.

What does it actually do better than the models I already know?

The short version: Fable handles longer, harder tasks with more reliability. Opus is excellent at most things. Where it occasionally struggles is on work that is genuinely long or complicated: a brief that builds on itself across many steps, a document set that requires holding many threads at once, code that depends on understanding a large codebase rather than a single function. Fable is substantially better at staying on track in those situations, catching its own mistakes, and completing the job without needing to be redirected.

Anthropic's benchmarks show it leading across software engineering, complex document analysis, vision tasks, and scientific reasoning. One useful frame from the people who have tested it: "default to Sonnet for 80 per cent of your daily work, escalate to Opus for hard tasks, reach for Fable when the assignment would otherwise warrant a senior contractor." That is a reasonable way to think about it.

There is also a fallback to know about. Fable includes safety classifiers that redirect certain types of request, mainly cybersecurity, biology, and chemistry topics, to Opus 4.8 instead. Anthropic says this affects fewer than 5 per cent of sessions, and you will see a notice when it happens. For most everyday uses it is invisible. If your work touches one of those fields, you may see more redirections than average, at least while Anthropic tightens the classifiers after launch.

What is this about tokens costing more?

A quick plain-English note on tokens, since this question comes up a lot with more powerful models. Tokens are roughly the unit AI systems use to measure how much text they read and write. More tokens means more processing, and more cost. When people say a model "uses a lot of tokens," they usually mean one of two things: the per-token price is higher, or the model tends to write very long responses.

For Fable, the concern is the first one. The API price is roughly double Opus: $10 per million input tokens and $50 per million output tokens. That is a real premium. The counterintuitive part is that Fable tends to complete equivalent work in fewer steps than Opus, so on complex multi-step jobs the actual cost gap is smaller than the per-token rate suggests. One company reported completing demanding spreadsheet work faster and with fewer turns than Opus; another said complex coding runs came in 25 to 30 per cent faster.

For simple tasks, short questions, or anything Opus or Sonnet already handles well, there is no good reason to use Fable. The per-token premium is pure overhead on straightforward work. Where the premium earns its keep is on genuinely difficult, long-running tasks where Fable's ability to hold context and work autonomously saves turns that would otherwise cost money anyway.

What about my privacy? I heard Anthropic now keeps the data.

This is the most important practical change in this launch, and it is worth taking seriously.

With every Claude model before Fable, business customers could request Zero Data Retention (ZDR), meaning Anthropic would not store your conversations at all. That agreement does not apply to Fable. Anthropic now mandates 30-day retention for all Fable traffic, on every platform where the model runs, with no exceptions. Your prompts and the model's responses are stored for 30 days. After that, they are deleted in almost all cases. Anthropic says it will not use this data to train new models, and has published controls: employees can only access flagged conversations through tooling that prevents copying or downloading, and every access is logged.

The reason Anthropic gives is legitimate. Some attacks on AI systems only become visible when you can look across many requests, not just one. Sophisticated jailbreak attempts, for example, may send hundreds of slight variations until one works. You cannot detect that pattern if you are only looking at one conversation at a time. The 30-day retention window gives Anthropic's safety team visibility across sessions.

The practical question for you is what you are putting into those sessions. For personal curiosity, research, or general writing, this change probably does not affect you much. For anything involving client information, health data, legal matters, financial details, or material your employer considers confidential, the calculus changes. Some large organisations have already blocked Fable access for their employees on this basis, exactly because their data-handling obligations do not permit 30-day retention on a third-party server, regardless of the privacy controls around it.

The honest summary: Fable is not suitable for confidential professional or regulated work until you have confirmed that your data-handling obligations allow it. For personal use, the change is worth knowing, but it is unlikely to affect most people's day-to-day experience.

When should I reach for Fable, and when should I not?

If you are on a Pro, Max, or Team subscription, Fable is included at no extra cost until 22 June 2026. After that date it will require usage credits on top of your subscription, and Anthropic has said it intends to restore it as a standard feature once capacity allows. This window is the right time to try it.

Tasks where Fable makes a noticeable difference: long documents that need sustained analysis across many sections; complex research that builds on itself across many steps; involved coding work in a large or unfamiliar codebase; anything where you have previously found yourself asking the same model the same question multiple times because it kept losing the thread.

Tasks where Fable is unlikely to feel different from Opus: short questions, one-shot writing tasks, quick summaries, emails, anything that Opus already handles in a single well-formed response. For those, Sonnet is usually the right choice anyway, and the extra capability is overhead you will not feel.

What I would avoid

I would avoid using Fable for anything confidential until you understand your data-handling obligations. The 30-day retention policy is not a flaw in Fable's design, it is part of how Anthropic is managing the safety risk of a more powerful model, but that does not make it neutral for sensitive work.

I would avoid switching to Fable as a default just because it is the most capable model available. The per-token cost is real, and for everyday work you will not notice the capability difference.

I would avoid assuming the response came from Fable if your topic touches cybersecurity, biology, or chemistry. You will see a notice when a classifier redirects to Opus, but if you are doing sustained work in those areas and the notice does not appear, that is still Opus answering, not Fable. The classifiers are intentionally broad while Anthropic refines them.

I would avoid paying for usage credits to continue using Fable after 22 June before you have confirmed, on your own actual work, that the quality difference is worth the cost. The free window before 22 June is exactly the right time to run that test.

A simple test before you pay for it

Take three to five pieces of work you have found frustrating with a previous model. Tasks where you needed more back-and-forth than felt reasonable. Tasks where the model seemed to lose track of something you had established earlier. Tasks where a long document needed sustained analysis rather than a surface summary. Feed each of them to Fable during the free window and compare the output with what you got before.

If the gap is obvious and the work is non-sensitive, Fable is worth keeping. If the gap is small or imperceptible for your actual work, save the usage credits. And if the work involves anything confidential, run that test with placeholder content, not the real material.

The verdict

Fable is a genuine step up in capability for difficult, long-running work. The per-token cost is higher, but it tends to finish complex jobs in fewer steps, so the real cost gap is smaller than the sticker price suggests. The free window until 22 June is the time to try it.

The thing to take seriously is the data. Fable is the first Claude model that does not support Zero Data Retention, and that matters if your work involves anything confidential. For personal use it is unlikely to change much. For professional or regulated work, check what your obligations are before you use it for anything real.