Claude Fable 5 is Anthropic’s first public Mythos-class AI model, launched on June 9, 2026, and positioned above Claude Opus. The short version: it’s meant for long, complex work in coding, research, vision, and knowledge tasks, while risky cybersecurity, biology/chemistry, and model-distillation prompts are automatically routed to Claude Opus 4.8 instead. That routing choice is the real story.
Claude Fable 5 goes public: what changed on June 9, 2026?
Anthropic has opened Claude Fable 5 to the general public as the first broadly available member of its new Mythos class. Until now, Mythos access was mostly something outsiders heard about rather than used: Anthropic unveiled Mythos Preview in April 2026, with availability limited to Project Glasswing partners.
The company describes Mythos as a tier above Claude Opus, not a routine model refresh. That matters because Opus has traditionally been Anthropic’s premium reasoning and agentic-work line; putting Mythos above it signals a new ceiling for long-horizon tasks.
Anthropic says Claude Fable 5 is state-of-the-art on nearly all tested benchmarks across software engineering, knowledge work, vision, and scientific research. The company’s more interesting claim is narrower: the longer and more complex the task, the larger Fable 5’s lead becomes. That’s exactly where frontier models have usually become expensive, forgetful, or erratic.
The Mythos-class idea, minus the marketing fog
Mythos-class is Anthropic’s label for its strongest model family as of 2026. Claude Fable 5 is the public-facing version, while Claude Mythos 5, launched the same day, is a more restricted version of the same underlying model with safeguards lifted in some areas.
That split is deliberate. Anthropic isn’t simply refusing to release its strongest system, and it isn’t throwing every capability onto the open consumer market either. It’s trying a third path: make the capable model broadly useful, then redirect specific high-risk use cases to a weaker model.
Honestly, that’s more interesting than another benchmark chart. A refusal-only policy frustrates normal users and gives little nuance. Full release creates obvious dual-use problems. Routing risky traffic to Claude Opus 4.8 is Anthropic’s attempt to keep the productivity gains of Claude Fable 5 while reducing access to the most sensitive capabilities.
How Claude Fable 5 handles risky prompts
When a user asks about high-risk cybersecurity, biology or chemistry, or model distillation, Anthropic says the request is automatically routed away from Claude Fable 5 and handled by Claude Opus 4.8. The user may still get help, but not from the Mythos-class model.
This is not a tiny footnote. It’s the core product decision. Instead of blocking the whole model behind government or enterprise gates, Anthropic is letting the public use Claude Fable 5 while putting a capability governor on the areas it sees as most dangerous.
The company says these safeguards trigger in less than 5% of sessions. That sounds low, but the lived experience will depend on your work. If you’re a software engineer asking about secure authentication, malware analysis in a defensive context, or red-team documentation, you may run into routing more often than a marketer drafting campaign copy.
A concrete way to read the figure: if your team runs 2,000 Claude sessions in June 2026, a 5% trigger rate would mean up to about 100 sessions could be routed to Claude Opus 4.8. At 1%, it would be 20. The difference is not academic if those sessions are the ones attached to your hardest security reviews.
Fable 5 vs Opus 4.8 vs Mythos 5
Anthropic’s naming now carries real policy consequences. Claude Fable 5 is the public Mythos-class model. Claude Opus 4.8 is the fallback for certain high-risk requests. Claude Mythos 5 is the restricted version aimed at approved organizations through Project Glasswing, including work with the US government.
For buyers, the practical question is not only “which model is smartest?” It’s “which model will actually answer the kind of work I do?” A public-interest biology researcher and a customer-support team could have very different experiences with the same subscription tier.
| Model | Availability in 2026 | Positioning | Key restriction |
|---|---|---|---|
| Claude Fable 5 | Public launch on June 9, 2026 | First public Mythos-class model, above Claude Opus | High-risk queries routed to Claude Opus 4.8 |
| Claude Opus 4.8 | Available as Anthropic’s Opus-tier model | Used as safer fallback for sensitive areas | Weaker than Mythos-class for the hardest tasks, by Anthropic’s framing |
| Claude Mythos 5 | Restricted to approved organizations via Project Glasswing | Same underlying model, safeguards lifted in some areas | Not generally public; tied to trusted access |
Project Glasswing partners include AWS, Apple, Google, Cisco, Microsoft, and JPMorgan Chase. That list tells you who Anthropic thinks needs early access to the sharper tool: cloud providers, platform companies, security vendors, and large financial infrastructure.
Anthropic describes Claude Mythos 5 as the strongest cybersecurity model in the world, while saying a trusted access program is planned. Treat that as a big claim rather than a settled independent fact until more outside evaluations land.
What can you actually use Claude Fable 5 for?
The best fit for Claude Fable 5 is work that gets worse when a model loses the thread. Long codebases. Dense research packets. Multi-document analysis. Scientific literature review. Visual reasoning over complex materials. Those are the places where Anthropic says the model’s lead grows with task length.
If you’re using Claude for ordinary short-form drafting, the difference may feel smaller. A three-paragraph email rarely justifies frontier-model drama. But give the model a messy repository, a 60-page technical brief, or a research question that requires holding many constraints at once, and the Mythos-class pitch starts to make sense.
Use Claude Fable 5 for:
- Reviewing large software changes and explaining architectural trade-offs.
- Summarizing long reports while preserving caveats and unresolved questions.
- Comparing scientific papers and extracting claims, methods, and limitations.
- Analyzing images or visual documents where context matters.
- Planning knowledge-work projects with many dependencies and constraints.
The pitfall nobody mentions enough: safety routing can make results inconsistent inside a single workflow. You may begin with Claude Fable 5 on a software design task, then get routed to Claude Opus 4.8 when the conversation touches exploit prevention or malware behavior. The model name may change less visibly than the quality of the answer.
Pricing, plans, and the data-retention catch
Pricing for Claude Fable 5 has been reported, but as of June 9, 2026 it appears to be single-source information rather than a fully settled public price sheet. Reported pricing is $10 per million input tokens and $50 per million output tokens. The same reporting says Claude Fable 5 is included in Pro, Max, Team, and Enterprise plans until June 22, 2026.
Here’s a quick calculation using those reported 2026 prices. A task with 200,000 input tokens and 20,000 output tokens would cost about $2 for input and $1 for output, or $3 total through the API. A heavier research run with 1 million input tokens and 100,000 output tokens would be about $10 plus $5, or $15.
At this price, if the reported numbers hold, it’s hard to do better for genuinely complex work. The catch is governance, not just cost. Anthropic now requires 30-day data retention on all Mythos-class traffic, including enterprises that previously had zero-retention agreements.
That requirement will stop some deployments cold. Regulated teams handling sensitive legal, healthcare, financial, or unreleased product data need to treat Claude Fable 5 differently from older enterprise Claude configurations. Don’t assume your existing zero-retention contract covers Mythos traffic. According to Anthropic’s 2026 policy, it does not.
Anthropic’s bigger bet: a brake pedal, not a locked door
Claude Fable 5 arrives while Anthropic is preparing an IPO and publicly urging major AI labs to adopt a coordinated “brake pedal” on frontier development. The stated concern is recursive self-improvement risk, where advanced systems could accelerate AI progress in ways that become hard to control.
That context colors the launch. Anthropic wants to show investors, governments, and enterprise buyers that it can commercialize its strongest work without pretending dual-use risks are imaginary. The answer it’s testing is model routing: strong model for most work, weaker model for risky domains, restricted access for the most capable version.
Is that enough? Maybe not. Automatic classifiers can overcatch harmless requests and miss cleverly phrased dangerous ones. Anthropic says external bug bounty testers spent more than 1,000 hours trying the system and found no universal jailbreak, which is reassuring, but not the same as proving the system is unbreakable.
The counter-argument is simple: routing risky prompts to a weaker model may create a false sense of safety if the weaker model is still highly capable. Claude Opus 4.8 is not a toy. For many dangerous tasks, “weaker than Mythos” could still mean “strong enough to matter.”
FAQ: Claude Fable 5
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s first Mythos-class AI model available to the general public, launched on June 9, 2026. It sits above the Claude Opus tier and is aimed at long, complex work in coding, research, vision, and knowledge tasks.
Is Claude Fable 5 the same as Claude Mythos 5?
No. Claude Fable 5 is the public Mythos-class model, while Claude Mythos 5 uses the same underlying model with safeguards lifted in some areas and is restricted to approved organizations through Project Glasswing.
Why does Claude Fable 5 route some prompts to Opus 4.8?
Anthropic routes high-risk cybersecurity, biology/chemistry, and model-distillation queries to Claude Opus 4.8 to reduce access to the most sensitive Mythos-class capabilities. The company says safeguards trigger in less than 5% of sessions and may catch harmless requests.
Does Claude Fable 5 keep enterprise data?
Yes, for Mythos-class traffic. Anthropic requires 30-day data retention on all Mythos-class usage in 2026, even for enterprises that previously had zero-retention agreements.
How much does Claude Fable 5 cost?
Reported 2026 pricing is $10 per million input tokens and $50 per million output tokens, but that figure should be treated as reported rather than fully confirmed. It has also been reported as included in Pro, Max, Team, and Enterprise plans until June 22, 2026.
EN 
FR 
ES