Claude Fable 5 — What Anthropic’s New Model Can Do and Whether It’s Worth Your Attention

This is not just another Claude update

Claude Fable 5, released by Anthropic on June 9, 2026, does not look like a routine family refresh. The company presents it as the public, safeguarded version of Mythos 5 — a new class of models positioned above the Opus line. In practice, that means one thing: users get more power for long, complex work, but without fully opening up the model’s highest-risk capabilities.

So if you want the simplest label, it is this: Claude Fable 5 is Anthropic’s most capable widely released model, but with hard safety rails. That matters, because this balance between capability and restriction is exactly what determines who Fable 5 is actually for.

What really makes Fable 5 different

Anthropic says Fable 5 is state of the art on nearly all tested AI capability benchmarks and that its lead grows as a task becomes longer, more complex, and more autonomous. This is not mainly a fast-chat model. It is a model for cases where AI has to plan, work over a large context, use tools, and deliver a result across several or many steps.

The launch materials keep returning to four areas: software engineering, knowledge work, vision, and long-context reasoning. Anthropic highlights examples such as large-scale codebase migrations, financial document analysis, rebuilding apps from screenshots, and working over huge context windows with persistent notes and memory.

Specs that matter in practice

On the technical side, quite a lot is public. claude-fable-5 has a 1M-token context window, 128k max output tokens, always-on adaptive thinking, the same tokenizer as Opus 4.8, and a knowledge cutoff of January 2026. For vision-heavy work it supports images up to 2576 px on the long edge, and the API can accept as many as 600 images or PDF pages in a single request.

At the same time, some things remain undisclosed. Anthropic has not publicly shared parameter count, full architecture, or training compute — the kind of detail you often get with open-weight models. That matters, because Fable 5 is described more through behavior and positioning than through a classic hardware-style spec sheet.

Where Fable 5 makes the most sense

The most natural use case is code. Anthropic gives a Stripe example where Fable 5 reportedly compressed months of engineering work into days, including a migration across a 50-million-line Ruby codebase that would otherwise have taken a whole team more than two months by hand. The company also emphasizes strong results on coding quality and long-horizon engineering benchmarks.

The second strong lane is analysis and documents. Anthropic points to finance benchmark results from Hebbia and partner feedback describing major gains in document reasoning, tables, charts, and multi-step analysis. If you compare models for research, investing, legal work, compliance, or business analysis, Fable 5 looks much more compelling than a model optimized mainly for quick answers.

The third lane is vision. Anthropic says Fable 5 can extract precise numbers from dense scientific figures and reconstruct web-app source code from screenshots. That is far more concrete than a vague claim like “the model can see images.”

The biggest strengths

Fable 5’s biggest advantage is not just “more intelligence.” It is better performance across longer horizons. Anthropic and its launch partners repeatedly highlight fewer turns, better autonomy, cleaner tool use, and stronger token efficiency. In practice, that can mean less manual nudging and a higher chance that the model delivers a useful first pass on its own.

The launch materials also contain several useful quality signals: breaking 90 percent on an internal benchmark for complex analytical tasks, stronger spreadsheet performance, and clear gains in app prototyping. Of course, part of this evidence comes from Anthropic or early-access partners, so it should be read as directional rather than absolute. Still, the message is consistent: Fable 5 is meant for real work, not just polished answers.

Limitations, privacy, and safety

This is where Fable 5 stops being universal. Anthropic built new safety classifiers around it. If a request touches cybersecurity, biology, chemistry, or attempts to extract reasoning capabilities at scale, the response may be handed off to Claude Opus 4.8 or ended with a refusal. The company also says the safeguards are intentionally conservative at launch, which means some harmless prompts will get caught too. Per Anthropic’s early data, the handoff occurs in fewer than 5% of sessions — more than 95% of the time, Fable 5 answers without switching.

On privacy, one detail matters a lot: Claude Fable 5 on the API requires 30-day data retention and is not available under zero data retention. That will be a real deployment constraint for some organizations. The better news is that Anthropic says, for commercial products, it does not train on your inputs and outputs by default, unless you explicitly opt in through a partner or training program. Consumer products work differently: users may allow model improvement, but incognito chats are not used to improve Claude.

Then there is price. Fable 5 costs twice as much as Opus 4.8 at the token level. So the right question is not simply “is it better?” but “is it better enough to justify the cost in my workflow?”

Fable 5 vs GPT-5.5, GPT-4.1, Gemini and Llama

A quick reference table first:

Model	Context / max output	Price (input / output per 1M tokens)	Bottom line
Claude Fable 5	1M / 128k	$10 / $50	the strongest public Claude for long work, with hard safeguards
GPT-5.5	1M / n/a	$5 / $30	OpenAI’s current flagship, omnimodal and cheaper
GPT-4.1	~1M / 32.8k	$2 / $8	the cheapest long context in the GPT family
GPT-4o	128k / 16.4k	$2.50 / $10	popular and cheap, but short context
Gemini 3.1 Pro	1M / 64k	$2 / $12 (under 200k tokens)	Google’s strongest current rival
Llama 4 Maverick	1M / —	open weights, cost depends on hosting	maximum deployment freedom

Start with OpenAI’s current flagship. GPT-5.5 (launched April 23, 2026) also offers a 1M-token context window and costs $5 / $30 per million tokens — half the input price and 40% less on output than Fable 5. It is the strongest direct rival today: natively omnimodal and the record-holder on the Artificial Analysis Intelligence Index. Anthropic’s answer is positioning: Fable 5 is the public version of a model class above Opus, tuned for the longest, most autonomous work — with safeguards GPT-5.5 does not carry in this form. (More background on that rivalry: our April 20–26 weekly recap.)

Against GPT-4o, Fable 5 mainly wins on context and long-horizon positioning, but loses on price. GPT-4o has a 128k context window and costs $2.50 / $10 per million tokens, so for many lighter or faster use cases it will still be the more economical option.

Against GPT-4.1, the comparison gets more interesting, because OpenAI also offers roughly 1M context at a lower price and a 32,768-token max output. If your evaluation is mostly about API cost and long context, GPT-4.1 is a very serious benchmark. Fable 5, in contrast, is being sold more as a “hardest agentic work” model and as the safe public version of a Mythos-class system.

On Google’s side, the relevant competitor is now Gemini 3.1 Pro, not PaLM. That is worth saying plainly, because plenty of search queries still mention PaLM — yet text-bison and chat-bison were officially retired on April 21, 2025. PaLM is now a historical category, not a real product choice.

And if you compare Fable 5 with Llama 4 Maverick, you are really comparing two different product philosophies. Fable 5 is managed, closed, and heavily safety-wrapped. Llama 4 Maverick offers a more flexible deployment story: open weights, downloads, Llama API, and partner hosting. It is a good illustration that the “best model” also depends on whether you care more about turnkey safety or deployment control.

Three prompts worth trying first

The examples below are tuned to the areas Anthropic emphasizes most: long coding tasks, document analysis, and vision. If the prompt drifts into offensive cybersecurity, biology, or chemistry, Fable 5 may switch to Opus 4.8 or refuse to answer.

Code refactor prompt

You are a staff engineer leading a migration across a large repository.
Goal: migrate from [old library] to [new library].
First:
1. Map every dependency on the old library.
2. Break the migration into phases with risk, rollback, and tests.
3. Identify which files can be changed automatically and which require manual review.
4. Return the result in this format: plan, file list, possible regressions, sample commits.
If anything is missing, ask no more than 5 precise questions.

Document analysis prompt

Review the attached documents and prepare a memo for the executive team.
I need:
- an 8-point executive summary,
- 3 biggest risks,
- 3 biggest opportunities,
- a table: claim / evidence / confidence level / missing data,
- a recommendation: act now, wait, or reject.
Do not summarize everything. Show only what changes the decision.

Vision prompt for screenshot-to-app

Based on the attached screenshot, reconstruct the interface structure.
First describe:
- layout,
- components,
- information hierarchy,
- styles worth preserving.
Then produce:
- an implementation plan,
- a React component structure,
- starter code,
- a list of ambiguities that need clarification.
Do not guess when an element is not visible — mark it as unspecified.

Verdict

Claude Fable 5 currently looks like one of the most interesting public AI launches of 2026, especially for people who need more than a fast conversational model. If you build software, work with large documents, run research workflows, or want to turn screenshots into usable prototypes, Fable 5 is a model you cannot ignore.

But it is not the perfect model for everyone. It is more expensive than Opus 4.8, it comes with stronger safety restrictions, and it requires 30-day retention on the API. So the most accurate conclusion is not “this is the new universal winner,” but rather: this is Anthropic’s strongest public Claude for long, complex work — as long as your use cases stay inside its safe operating boundaries.

FAQ

Is Claude Fable 5 the same as Claude Mythos 5?

No. Anthropic says they are the same underlying model, but Fable 5 adds safeguards, while Mythos 5 is restricted to the limited trusted-access program (Project Glasswing).

Is Fable 5 better than Opus 4.8?

Anthropic positions Fable 5 as its most capable widely released model, and the migration guide treats it as the upgrade path for the hardest workloads. At the same time, Opus 4.8 remains cheaper and acts as the fallback for requests covered by Fable 5’s safeguards.

Can I use Fable 5 for cybersecurity or biology work?

Only in a very limited sense in the public version. Anthropic routes cyber, bio, and chemistry requests to Opus 4.8 or refuses them, because those domains were judged too risky for broad public deployment.

How much does Fable 5 cost and where is it available?

Public pricing is $10 / $50 per 1M input / output tokens. The model is generally available on the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry; on paid Claude subscriptions it is included at no extra charge through June 22, 2026, and from June 23 it is expected to move to usage credits unless Anthropic extends the window.

Is Fable 5 suitable for sensitive data?

For commercial use Anthropic says it does not train on inputs and outputs by default, but Fable 5 on the API still requires 30-day retention and does not support zero data retention. If privacy is critical, that requirement needs to be reviewed before deployment.

Sources: Anthropic (anthropic.com) — official Claude Fable 5 and Mythos 5 announcement, model documentation, migration guide and data-retention policies. Press: Reuters, The Verge, Ars Technica, TechCrunch. As of June 9, 2026.