Anthropic Releases Claude Fable 5, Its First Public Mythos-Class Model
Anthropic just released Claude Fable 5, the first publicly available Mythos-class model and the first mainstream AI system designed to sometimes hand conversations to another model mid-session.
Fable 5 is the public-facing version of Mythos 5, Anthropic’s highest capability tier above the Claude Opus line. Until now, Mythos-class systems had remained restricted to limited-access programs, including Project Glasswing, where Anthropic provided early access to cybersecurity and critical-infrastructure partners.
What makes Claude Fable 5 different is not a separate architecture, but a deployment layer built around routing and safeguards. On most requests, users interact directly with the same underlying frontier model Anthropic previously declined to release broadly. On a smaller subset of conversations, however, the system automatically switches the interaction to Claude Opus 4.8 instead.
That design choice may mark a turning point in how frontier AI systems are deployed publicly. Fable 5 is one of the first consumer AI products where safety is implemented not primarily through refusals, but through live model substitution.
What Claude Fable 5 Can Do
Anthropic and several early testers describe substantial capability gains over previous Claude 5 models, especially on long-running and multi-step tasks.
Stripe reported that Claude Fable 5 completed a migration across a 50-million-line Ruby codebase in a single day, work the company estimated would have taken engineers months manually. Analytics platform Hex said Fable was the first model to surpass 90% on its benchmark for complex, long-duration workflows.
Anthropic also highlighted improvements in vision and autonomous execution. The company says Fable 5 can reconstruct web application source code directly from screenshots and completed Pokémon FireRed using raw screen-frame input without the navigation scaffolding earlier Claude systems required.
The central theme across those demonstrations is endurance rather than isolated benchmark spikes. Anthropic says the performance gap between Fable 5 and earlier Claude models widens as workflows become longer, more iterative, and more agentic.
On short prompts, the difference may appear incremental. On workflows spanning hours or days, the model behaves meaningfully differently.
Claude Fable 5 and its restricted sibling, Mythos 5, are priced at $10 per million input tokens and $50 per million output tokens.
How Claude Model Switching Works
The defining feature of Claude Fable 5 is its automated routing system.
Every request is evaluated by classifiers that determine whether the conversation falls into a high-risk category. If the system flags the interaction, Fable 5 does not continue responding. Instead, the conversation is rerouted to Claude Opus 4.8 inside the same thread.
The user receives a notice that the model switched, and subsequent responses are labeled according to the model currently replying.
Anthropic says the routing system focuses on three categories:
- Offensive cybersecurity activity
- Biology and chemistry workflows
- Model distillation or capability extraction attempts
The company’s reasoning is straightforward: the same capabilities that accelerate legitimate scientific or technical work could also materially assist cyberattacks or biological misuse.
Rather than withholding the model entirely, Anthropic chose a layered deployment strategy. Most conversations run on the Mythos-class system directly, while higher-risk requests fall back to a more heavily constrained model with established safeguards.
Anthropic also notes that the classifiers evaluate everything the model can access, not just the user’s latest message. Uploaded files, memory context, connector data, and web search results can all contribute to triggering a switch.
That means a conversation may reroute because of contextual information the user never explicitly typed.
Why Claude Opus 4.8 Fallback Exists
The Claude Opus 4.8 fallback mechanism reflects a broader shift in how frontier AI systems are being operationalized.
Historically, AI safety layers primarily relied on refusals. A model either answered or declined. Claude Fable 5 introduces a different approach: dynamic orchestration between models with different capability and risk profiles.
In practice, this turns model identity into a runtime variable rather than a fixed user selection.
Anthropic says automatic switching is enabled by default across Claude’s consumer applications and workplace integrations, including Microsoft 365, Teams, Slack, and Anthropic’s desktop and mobile apps.
After a switch occurs, conversations remain on Opus 4.8 unless the user manually returns to Fable 5. Developers using the API must configure fallback behavior themselves rather than relying on automatic routing.
For teams deploying Claude 5 systems in production, that creates a new operational consideration: applications may need to track not only outputs, but which model actually generated them.
Organizations working near cybersecurity, chemistry, biology, or sensitive research domains should expect occasional routing events and account for them in logging, auditing, and workflow design.
Anthropic has also signaled that future Mythos-class deployments will likely follow the same pattern: broad public access combined with dynamic fallback layers and separate trusted-access programs for vetted researchers.
Claude Fable 5 is arguably the most capable AI model Anthropic has released publicly. It also introduces something new to mainstream AI products: the possibility that the system answering your question may change while the conversation is still happening.
