What Is Claude Fable 5? Anthropic’s First Public Mythos-Class Model, Explained

Claude Fable 5 is the most capable AI model Anthropic has ever released to the public. It launched on June 9, 2026, and it is the first generally available model from the company’s “Mythos” class, a tier of models Anthropic had previously restricted to vetted partners because of safety concerns. Fable 5 is the same underlying model as the restricted Claude Mythos 5, wrapped in safeguards that block a small set of high-risk topics and quietly hand those requests to Claude Opus 4.8 instead. It costs double what Opus 4.8 costs on the API, it is temporarily included at no extra cost on paid Claude plans, and every conversation with it is retained for 30 days as a safety measure. This article explains what Fable 5 actually is, what it is genuinely good at, what the guardrails mean in practice, and who should bother switching to it.

I run CriticNest and hey-ash.com, and I have spent six years evaluating the software I write about before I write about it. I also run AI models in production every working day, so this is not a secondhand summary. The facts in this article were verified on June 10, 2026 against Anthropic’s official launch announcement, its model documentation, and same-day press coverage. Where a detail could not be verified from a primary source, it is not in this article.

What Is Claude Fable 5?

Claude Fable 5 is a new flagship model from Anthropic, the company behind the Claude chatbot and Claude Code. It sits above the Opus tier that has topped Anthropic’s lineup for the past two years. In the company’s own words from the launch announcement, Fable 5’s capabilities “exceed those of any model we’ve ever made generally available,” with state-of-the-art results on nearly every benchmark Anthropic tested, and the strongest gains in software engineering, knowledge work, vision, and scientific research.

The name that matters in that sentence is not Fable. It is Mythos. Fable 5 is a Mythos-class model, which means it is built on the same underlying system as Claude Mythos 5, a model Anthropic considers powerful enough that it does not sell it openly. Fable 5 is that system with a safety layer bolted on: requests in a few high-risk domains get blocked and silently answered by the older Claude Opus 4.8 instead. Everything else, which Anthropic says is more than 95 percent of real sessions, runs on the new model directly.

One pattern from the launch benchmarks is worth pulling out, because it tells you who this model is for: the longer and more complex the task, the larger Fable 5’s lead over previous models. For a quick question or a short draft, you will struggle to tell it apart from Opus 4.8. Hand it a multi-hour autonomous job, and the gap shows.

Mythos, Explained: The Model Anthropic Kept Locked Up

Mythos first surfaced in April 2026 as a preview available only to a limited set of partners, with Anthropic citing cybersecurity concerns as the reason for the restriction. The week before Fable 5 launched, that preview program expanded to roughly 150 organizations across more than 15 countries. The public framing was unusual for an AI company: Anthropic was openly saying it had built something it did not consider safe to hand to everyone, days before handing a safeguarded version to everyone.

The unrestricted version, Claude Mythos 5, ships through a program called Project Glasswing, a collaboration with the US government. Glasswing gives cybersecurity defenders and critical infrastructure providers access to the model with the cyber safeguards lifted, and a separate track gives select biology researchers access with the biology and chemistry safeguards lifted. Anthropic describes Mythos 5 as having “the strongest cybersecurity capabilities of any model in the world,” which is precisely why ordinary customers do not get that version.

So the lineup now reads like this: Mythos 5 is the raw model for vetted defenders and researchers. Fable 5 is the same model with the dangerous edges filed off, for everyone else. If you have followed AI long enough to be skeptical of safety theater, the honest read is that this is a real capability tier with a real gate in front of it, not a marketing exercise: the gate has a government program, a bug bounty, and a mandatory data retention policy attached to it, all of which cost Anthropic money and goodwill to run.

What Fable 5 Is Actually Good At

Launch-day benchmarks always deserve suspicion, so I am limiting this table to specific, attributable claims from the announcement, several of which come from named third-party companies rather than Anthropic itself:

The theme across all of these is endurance: long codebases, long documents, long research sessions, long games. If your use of AI is “summarize this email,” none of this matters to you. If you run agents that work for hours, it is the whole story. For context on what the previous generation of coding assistants could and could not do, my reviews of Cursor and GitHub Copilot cover the tier of tools that models like this one now power.

Fable 5 vs Opus 4.8: Pricing, and When Each Makes Sense

Anthropic’s API pricing as of June 10, 2026, per its model documentation:

Two times the price of Opus 4.8 is a real bill on heavy workloads, and Anthropic notes the new pricing is still less than half what the Mythos Preview cost its partners. The decision framework I would actually use:

• Pick Fable 5 when the task is long, autonomous, and expensive to get wrong: large refactors and migrations, overnight agent runs, deep research synthesis, complex document analysis. The launch data consistently shows it finishing faster with fewer wasted tokens, which claws back part of the price gap on its own.
• Stay on Opus 4.8 for routine production work, short interactive sessions, and anything where the older model already performs at ceiling. Anthropic itself uses Opus 4.8 as the fallback inside Fable 5, which tells you it considers it a fully trustworthy workhorse.
• Stay on Sonnet 4.6 or Haiku for high-volume, low-complexity calls. Nothing about this launch changes the economics of classification, extraction, or short-form generation.

The Safety Guardrails, and Why Your Answer Might Come From Opus 4.8

Fable 5 ships with safeguards in three domains: cybersecurity (exploitation and offensive tasks), biology and chemistry (weapons-adjacent and dual-use research), and distillation (attempts to extract the model’s capabilities to train a competitor). When a request lands in one of those areas, Fable 5 does not answer it. The request falls back to Claude Opus 4.8, which responds under its normal rules.

In practice, Anthropic says the fallback triggers in fewer than 5 percent of sessions, and early data shows at least 95 percent of Fable sessions run entirely on Fable’s own responses. If you are a developer, a writer, an analyst, or basically anyone not probing malware techniques or pathogen synthesis, you will likely never notice the seam.

The robustness numbers are stronger than the usual launch-day boilerplate: an external bug bounty logged more than 1,000 hours of testing without finding a universal jailbreak, the model complied with zero harmful single-turn requests across 30 published jailbreak techniques, and one external partner rated its cybersecurity safeguards the most robust of any model it had tested. The honest caveat, which Anthropic includes itself: the UK AI Security Institute made progress toward one jailbreak within its initial testing window. No guardrail in this industry has stayed unbroken forever, and this one will be attacked harder than most.

The 30-Day Retention Rule You Should Know About

How to Try Claude Fable 5

The practical takeaway sits in the third row: if you already pay for any Claude plan, you have a free window to test the most capable public model in the world until June 22. There is no reason not to use it. Run your hardest real task through it, compare against what Opus 4.8 gives you, and decide with your own evidence whether the 2x API price ever becomes worth paying.

Should You Care? The Honest Take

If you use AI casually, through the chat interface, for drafting and questions and everyday help: no urgency. Fable 5 will feel like a slightly sharper Claude, and the free window makes it worth a try, but Opus 4.8 was not the thing limiting you.

If you build with the API or run agentic workloads, this launch matters for a specific reason: the capability jump is concentrated exactly where agent budgets bleed, in long tasks that previous models needed retries, supervision, and wasted tokens to finish. A model that completes a two-month migration in a day, or matches four days of a rival’s research output in 36 hours on a third of the tokens, changes project math even at double the per-token price. The right move is not to switch everything; it is to identify your single longest, most failure-prone workload and benchmark Fable 5 on that.

And if you work in search or publishing, as I do, file this launch under the same trend I documented in my breakdown of Google’s AI search guidance: the models reading, summarizing, and citing the web are getting dramatically better at long-document comprehension. The bar for content that survives AI-mediated search keeps rising, and launches like this one are why.

Frequently Asked Questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic’s most capable publicly available AI model, released on June 9, 2026. It is the first generally available model from Anthropic’s Mythos class, with safeguards that block high-risk topics and route those requests to Claude Opus 4.8 instead. Anthropic reports state-of-the-art results in software engineering, knowledge work, vision, and scientific research.

What is Claude Mythos?

Mythos is Anthropic’s restricted top model tier. It launched as a limited preview in April 2026 over cybersecurity concerns, and Claude Mythos 5, the unrestricted version of Fable 5, is available only to vetted cyberdefense organizations through Project Glasswing, a collaboration with the US government, plus select biology researchers. Anthropic describes it as having the strongest cybersecurity capabilities of any model in the world.

How much does Claude Fable 5 cost?

On the Claude API, Fable 5 costs $10 per million input tokens and $50 per million output tokens, double the price of Claude Opus 4.8, which costs $5 and $25. Anthropic notes this is less than half the price its partners paid for the earlier Mythos Preview.

Is Claude Fable 5 free with a Claude subscription?

Temporarily, yes. Pro, Max, Team, and seat-based Enterprise plans include Fable 5 at no extra cost from June 9 through June 22, 2026. From June 23 it requires usage credits, and Anthropic says it intends to restore it as a standard plan feature when capacity allows.

How is Fable 5 different from Claude Opus 4.8?

Fable 5 is a newer, more capable model class. Anthropic’s launch data shows the gap widening as tasks get longer and more complex: it beat Opus 4.8 on spreadsheet benchmarks at every effort level while finishing 25 to 30 percent faster, jumped 10 points past Opus on a core analytics benchmark, and reached the final act of a long strategy game three times more often in memory testing. For short, simple tasks the two are hard to tell apart.

Why do some Fable 5 answers come from Opus 4.8?

Fable 5 ships with safeguards covering cybersecurity exploitation, biology and chemistry risks, and attempts to distill the model’s capabilities. Requests in those areas are answered by Claude Opus 4.8 instead. Anthropic reports the fallback triggers in fewer than 5 percent of sessions.

Is Claude Fable 5 safe?

Anthropic put unusual effort into the launch safeguards: an external bug bounty found no universal jailbreak in more than 1,000 hours of testing, and the model complied with zero harmful single-turn requests across 30 published jailbreak techniques. The honest caveat is that the UK AI Security Institute made progress toward one jailbreak during its initial testing window, and no AI guardrail has stayed unbroken indefinitely.

Does Anthropic store my Fable 5 conversations?

Yes. All Mythos-class traffic, including Fable 5, is subject to a mandatory 30-day retention policy. Anthropic states the data is used only for safety monitoring, is not used for training, has logged human access, and is deleted after 30 days in almost all cases. Users with strict data retention requirements should review their agreements before routing sensitive work through Fable 5.

Who should use Claude Fable 5?

The model is built for long, complex, autonomous work: large codebase migrations, multi-hour agent runs, deep research, and dense document analysis. Casual chat users will notice little difference from Opus 4.8. API users with long-running agentic workloads are the audience the launch benchmarks speak to, and the strongest results, like Stripe’s one-day 50-million-line migration, come from exactly that kind of work.

What is Project Glasswing?

Project Glasswing is Anthropic’s collaboration with the US government that distributes the unrestricted Claude Mythos models to cybersecurity defenders and critical infrastructure providers. The Mythos preview program expanded to roughly 150 organizations across more than 15 countries in the week before Fable 5’s public launch.

Bottom Line

Claude Fable 5 is a genuinely new capability tier delivered with an unusually candid safety story: Anthropic built something it considers too dangerous to ship raw, said so publicly, and shipped a safeguarded version at double the old flagship price while keeping the raw version behind a government program. The benchmarks that matter are the endurance ones, and they are strong enough that anyone running serious agentic workloads should benchmark it before June 22, while the subscription window makes that free. Everyone else can relax: Opus 4.8 did not stop being excellent on June 9, and it now has the strange honor of being the safety net inside its own successor.

About the Author

Ashikur Rahman is the founder of CriticNest and hey-ash.com, where he has spent six years as a solo operator building, auditing, and ranking content properties, and where AI models do real production work every day. Every article on CriticNest follows the same rule: claims are verified against primary sources on the day of publication, and what cannot be verified does not get published.