On June 9, 2026, Anthropic released two new versions of its Claude AI model — Claude Fable 5 and Claude Mythos 5. Both are built from the same underlying technology but tuned for different purposes. Anthropic

The biggest change with both models is that they can now work on their own for much longer periods than before. Think of it like the difference between asking an assistant to answer a single question versus handing them a multi-week project and checking in at major milestones rather than after each task.

Fable 5: For Long Projects

Claude Fable 5 is designed for work that stretches over time. Earlier versions of Claude were built for quick tasks — writing a piece of code, summarizing a document, or a back-and-forth conversation. Fable 5 is different. It's meant to hold onto a goal, break down large projects into steps, and work through them without needing constant human input.

This matters for engineering teams and researchers. If you need an AI to handle a full software development cycle, analyze a mountain of research papers, or run a data pipeline that takes hours, Fable 5 is what you'd reach for. Anthropic says Fable 5 has the same raw power as its top-tier models — it just has the stamina to match bigger ambitions.

Mythos 5: For Specialized Work

Claude Mythos 5 is an upgrade focused on three specific areas: cybersecurity, biology, and healthcare. Anthropic tested the new model on problems in those fields and found it performs better than the previous version.

These three areas matter because they're heavily regulated and involve real stakes — security, disease prevention, and treatment decisions. If doctors, researchers, or security teams use AI tools in their work, those teams need to know the model works reliably in those domains. Mythos 5 also gets the same extended working duration as Fable 5, which helps when you're doing the kind of complex, multi-step reasoning those fields require.

One Model, Two Flavors

Anthropic chose to release two versions from a single core model rather than building two entirely separate models from scratch. This is becoming common in AI labs. Instead of training multiple models for different jobs, they adjust how the same model behaves — like tuning an instrument for different purposes. One tuning emphasizes the ability to work alone for longer. The other emphasizes expertise in specialized domains. This approach saves time and computing power.

For people building products with Claude, the question becomes practical: which version do I need. Teams building automated systems that run for hours or days with minimal human checking should use Fable 5. Teams building tools for security, healthcare, or research should use Mythos 5. Some projects might need both, though pricing and how Anthropic handles that overlap remains to be seen.

Autonomy and Trust

Both models can work longer without a human checking in. This is not just a headline feature — it changes how teams need to think about these systems.

When an AI system can make decisions and take actions for hours without human oversight, one mistake can compound. If the system makes a wrong choice midway through a task, that wrong choice has more room to spread before anyone catches it. Teams using these models will need to think carefully about where to place checkpoints — moments when a human verifies the work is on track — even if the AI doesn't technically require them.

This is a familiar tension in technology. When database automation became practical in the 1990s, companies had the tools to automate business workflows, but many hadn't built the oversight and governance structures to manage them safely. AI agents are heading down a similar path. The capability is real, the use cases are clear, but the question of how to govern and audit these systems at scale is still being answered.

Testing and Proof

Anthropic has released benchmark data showing where Mythos 5 improves on its predecessor in those three specialized fields. The detailed technical documentation is available alongside the announcement. However, independent testing by outside researchers will be the next important check. Different AI labs test models differently, so direct comparisons can be tricky without using the same test set.

For companies considering using these models in regulated fields like healthcare or security, the benchmarks are a useful signal, but they shouldn't be the only factor. The right approach is to test the model on your actual work — with experts in your field reviewing the results — to be sure it works for your specific needs.

What's Next

Both models are available now through Anthropic's standard developer channels. Teams already using earlier Claude models for automation and extended tasks are the first audience. The broader direction Anthropic is charting suggests that the line between "a model" and "an automated worker" keeps blurring. Fable 5 in particular looks less like a conversational AI and more like a purpose-built tool for autonomously completing work. Whether that holds up when thousands of teams use it simultaneously at real scale is what we'll find out over the coming months.