Anthropic has released a lot of models over the past two years, and most of them slotted neatly into the lineup: a bit faster here, a bit smarter there. Claude Fable 5 is not that. It’s the first widely released model from Anthropic’s new top tier — internally called Mythos-class — and it was built for work that runs for hours or days rather than seconds.

It also had one of the messiest launches in recent AI history, including a three-week period where nobody could use it at all. If you’ve been confused about what Fable 5 actually is, whether you have access to it, and what it’s good for, this guide covers all of it.

What is Claude Fable 5?

Claude Fable 5 launched on June 9, 2026 as Anthropic’s most capable generally available model. It sits above Claude Opus 4.8 in the lineup, and it was designed around a specific idea: instead of answering a question well, it should be able to take on a whole project and see it through.

There’s a companion model, Claude Mythos 5. Same underlying model, same capabilities — but with several safety restrictions removed. Mythos 5 isn’t public. It’s deployed to a small set of vetted organizations through Project Glasswing, mostly cyber defenders and infrastructure providers working with the US government.

So the simplest way to think about it: Fable 5 is Mythos 5 with guardrails, made safe enough for general release.

The launch timeline, briefly

This matters because you may have run into a wall at some point and wondered what happened.

• June 9, 2026 — Fable 5 and Mythos 5 launch.
• June 12, 2026 — Anthropic suspends access to both models to comply with US Department of Commerce export controls.
• July 1, 2026 — The controls are lifted and access is restored, with upgraded cybersecurity safeguards.
• July 20, 2026 — Plan access is finalized after three separate extensions.

The instability is over. Terms are now permanent, which is the part most teams were waiting on before committing.

Key features of Claude Fable 5

1. Long-horizon autonomy

This is the headline capability. In an agent harness like Claude Code or Claude Managed Agents, Fable 5 can work for days at a time — planning across stages, delegating to sub-agents, and checking its own output as it goes. Prior models tended to drift or lose the thread on tasks that long. Fable 5 holds instructions across extended runs.

Zapier reported that where Opus 4.8 would stop and ask, Fable 5 keeps going. That’s the practical difference: fewer interruptions per unit of work.

2. First-shot correctness

Several early testers reported single-pass implementations of systems that used to take days of back-and-forth. Gusto said Fable 5 handled work in one day, inside a 50-million-line Ruby codebase, that would have taken their team more than two months by hand.

Your mileage will vary with how well-specified the task is. Vague prompts still produce vague results. But on complex, clearly defined problems, the first output is far more likely to be the final output.

3. Genuinely useful vision

Fable 5 reads diagrams, charts, and tables buried inside files and PDFs with substantially better accuracy than earlier models — often using fewer output tokens to do it. It’s trained to use bash and crop tools when an image is flipped, blurry, or noisy.

The knock-on effect for coding is interesting: the model uses vision to check its own front-end work against the original design, rather than just assuming the code it wrote produced the right result.

4. A 1 million token context window

Fable 5 ships with a 1M token context window by default and supports up to 128k output tokens per request. Large codebases, long document sets, and multi-hour agent sessions fit without aggressive trimming.

5. Better delegation

Fable 5 dispatches parallel sub-agents more readily and more reliably than previous Claude models, and it manages ongoing communication with long-running sub-agents rather than blocking until each one returns. If you’re building multi-agent systems, this is the change that most affects your architecture.

6. Professional-grade knowledge work

Not just code. In blind reviews at one legal tech company, lawyers found Fable 5’s contract redlines matched or beat their existing model every time. On finance benchmarks, testers reported double-digit gains in document reasoning and chart interpretation. Spreadsheets, slides, and long documents all improved.

What’s new compared to Claude Opus 4.8

If you’re migrating, these are the changes that will actually affect you.

Adaptive thinking is always on. It’s the only thinking mode available on Fable 5. You can’t disable it, and there are no extended thinking budgets. Depth is controlled through the effort parameter instead.

Raw reasoning is never returned. You can request a readable summary of the model’s reasoning (thinking.display: "summarized") or get empty thinking blocks ("omitted", the default). The raw chain of thought isn’t available on Fable 5 at all.

Requests can be refused by design. Fable 5 runs safety classifiers covering offensive cybersecurity, biology and life sciences, and attempts to extract its summarized reasoning. When one fires, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response — not an error. Your integration needs to handle that case.

Turns are much longer. A single request on a hard task at high effort can run for many minutes. Autonomous runs can go for hours. Anthropic explicitly recommends adjusting client timeouts, streaming, and progress indicators before migrating, and restructuring harnesses to check on runs asynchronously rather than blocking.

Safeguards affect consumer apps too. In Claude.ai and the Claude apps, flagged queries route automatically to Opus 4.8. You’ll see a notice that the model switched, and you’re not charged Fable prices for the rerouted response. Anthropic has said these safeguards are tuned conservatively and trigger in under 5% of sessions on average — so benign requests occasionally get caught.

30-day data retention is mandatory. Fable 5 and Mythos 5 are designated Covered Models for safety monitoring. Zero data retention is not available on either. This is a real blocker for some regulated workloads, so check it before you plan a migration.

Pricing and plan access

On Claude subscriptions

Access was finalized on July 20, 2026, and it splits by plan:

Fable 5 draws on your limits faster than other Claude models, which is why the 50% cap exists. When you hit it, you can either buy usage credits or switch to Sonnet 5 or Opus 4.8 for the rest of the week.

On the API

Fable 5 costs $10 per million input tokens and $50 per million output tokens, with the existing 90% input discount for prompt caching. That’s roughly double Opus 4.8’s rate, so model routing matters — don’t send bulk or routine work to Fable.

US-only inference is available at 1.1x pricing if you need workloads to stay in-region.

You are not billed for a request that’s refused before any output is generated, and fallback credit refunds the prompt-cache cost when you retry on another model, so you don’t pay that twice.

Where you can access it

Claude.ai and the Claude apps, Claude Code (CLI, web, and desktop), Claude Cowork, the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.

How to use Claude Fable 5

In Claude.ai or the Claude app

1. Open Claude on web, desktop, or mobile and sign in to a paid plan.
2. Tap the model selector at the top of the chat.
3. Choose Claude Fable 5.
4. Start your task.

That’s the whole setup. The more useful question is what to send it, which brings us to the part most people get wrong.

Start at the top of your difficulty range

Anthropic’s own guidance is blunt about this: teams seeing the best results point Fable 5 at their hardest unsolved problems, and testing it only on simple tasks undersells what it can do. Give it something harder than you’d give Opus. Ask it to scope the work, ask clarifying questions, then execute end to end.

Explain why, not just what

Fable 5 performs measurably better when it understands intent. Context lets it connect your request to relevant information instead of guessing at what you meant. A useful template:

I’m working on [the larger task] for [who it’s for]. They need [what the output enables]. With that in mind: [request].

Ask for the outcome first

At higher effort, Fable 5 can over-explain — surveying options it won’t pursue, narrating root causes at length. A short instruction fixes it better than a long list of rules:

Lead with the outcome. Your first sentence should answer “what happened” or “what did you find.” Supporting detail comes after.

Tell it not to gold-plate

Left alone at high effort, Fable 5 sometimes tidies up code you didn’t ask it to touch. Try:

Don’t add features, refactor, or introduce abstractions beyond what the task requires. Do the simplest thing that works well.

Give it a memory file

Fable 5 does noticeably well when it can record lessons from previous runs and reference them later. A plain Markdown file is enough — one lesson per entry, a one-line summary at the top, updated rather than duplicated.

For developers: the API basics

The model ID is claude-fable-5. Use the effort parameter as your main control:

• high — the default for most work
• xhigh — the most capability-sensitive tasks
• medium / low — routine work

Worth knowing: Fable 5 at low effort often outperforms previous models at their maximum. Reaching for xhigh on everything wastes money and time.

Supported features at launch include the memory tool, code execution, programmatic tool calling, compaction, vision, task budgets (beta), and tool-result clearing through context editing (beta).

One migration trap: if your existing prompts or Skills instruct Claude to explain or reproduce its internal reasoning in its response, those instructions can trigger the reasoning_extraction refusal category and cause elevated fallbacks to Opus 4.8. Audit your system prompts for show-your-thinking language before you migrate. Read the structured thinking blocks instead.

Long autonomous runs

Two instructions worth adding to any pipeline where nobody is watching in real time:

Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for; if something is not yet verified, say so explicitly.

You are operating autonomously. The user cannot answer questions mid-task. For reversible actions that follow from the original request, proceed without asking.

Anthropic reports the first one nearly eliminated fabricated status reports in testing, even on tasks specifically designed to provoke them.

Should you use Fable 5 for everything?

No, and Anthropic doesn’t suggest you should. At $10/$50 per million tokens and multi-minute response times, Fable 5 is the wrong tool for drafting an email or answering a quick question. Sonnet 5 and Haiku 4.5 exist for a reason.

The rough test: if the task is something you’d expect a capable person to spend a full day on or longer, and you’d normally have to babysit an AI through it, Fable 5 is worth the cost. If it’s something you’d finish in ten minutes, it isn’t.

Frequently asked questions

Is Claude Fable 5 available on the free plan? No. It’s available on Pro, Max, Team, and Enterprise plans only.

What’s the difference between Fable 5 and Mythos 5? They’re the same underlying model. Mythos 5 has safety classifiers for cybersecurity and biology removed, and it’s only available through Project Glasswing to a small group of vetted organizations.

Why did Claude switch to Opus 4.8 mid-conversation? A safeguard classifier flagged your query, most likely as cybersecurity- or biology-related. The routing is automatic, and you aren’t charged Fable prices for that response. False positives do happen — Anthropic tuned the classifiers conservatively.

How big is Fable 5’s context window? 1 million tokens by default, with up to 128k output tokens per request.

Can I use Fable 5 with zero data retention? No. Fable 5 and Mythos 5 both require 30-day data retention for safety monitoring and are not available under ZDR.

The bottom line

Fable 5 is a different kind of tool from what came before it, and using it like a faster chatbot will leave most of its value on the table. The models it replaces were assistants. This one is closer to a colleague you hand a project to and check in on tomorrow.

If you’re on Max or Team Premium, it’s already sitting in your model picker. Point it at the thing you’ve been putting off because it was too big.