Anthropic releases Mythos-level model for general use
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Sarah Grillo/Axios
Anthropic is releasing a safer, general-use Mythos-class model that exceeds the capabilities of all its prior releases, the company says.
Why it matters: Anthropic went from deeming this model so disruptive that no existing safeguards were sufficient and only a handful of organizations could access it, to releasing a public version less than three months later.
Driving the news: Anthropic unveiled Fable 5 today, its first Mythos-class model for general use — with safeguards that prevent everyday coders from using the model to hack infrastructure or ask about sensitive biological capabilities.
- Anthropic is also launching Claude Mythos 5 for users who already had access to Claude Mythos Preview, including partners in Project Glasswing.
- Anthropic will lift some safeguards for Claude Mythos 5 and deploy the model in collaboration with the U.S. Government, Anthropic said.
- The model will be available to a small group of cyberdefenders and infrastructure providers.
- The model is better at knowledge work, software engineering, scientific research and more, outscoring competing models from OpenAI and Google DeepMind on all benchmarks, per Anthropic.
How it works: Fable 5 will launch with safeguards that automatically re-route higher risk queries to lower performing models.
- Rather than answering certain cyber, biology, chemistry and model-distillation requests directly, Fable 5 automatically hands them off to the less-capable Claude Opus 4.8 model.
- Opus 4.8 doesn't have the reasoning capabilities to complete these high-risk requests.
- The underlying model for Mythos and Fable 5 is the same.
What they're saying: Dianne Penn, Anthropic's head of product management, research and labs, told Axios that Anthropic ran Fable 5 through internal safety evaluations, internal red-teaming and external adversarial testing before release to ensure safety.
- Penn said Anthropic is being "deliberately more conservative" at launch, meaning some legitimate scientific or security work could also be routed to Opus 4.8.
- Anthropic expects those fallbacks for benign requests to decline over time as testing continues after launch.
Follow the money: Fable 5 is double the price of Anthropic's Opus models, making it the most expensive release yet.
- Anthropic stresses that expense is relative: the Mythos-class model also gives you more intelligence and better performance, giving customers a "lower overall cost per task," Penn said.
Flashback: When Anthropic first previewed its Mythos-class models, the company said broader access would depend on whether it could build sufficient safeguards.
- In recent testing, Anthropic found that Mythos just needed 31 minutes to write an exploit for an already disclosed vulnerability in the Windows kernel.
What to watch: Anthropic is working on a more standardized trusted-access program to expand availability of Mythos 5 to vetted users, but Penn said there is no clear timeline.
- OpenAI has established a similar model for rolling out its cyber-capable models, including giving unrestricted access only to vetted security researchers, government partners and corporate partners.
The bottom line: Anthropic believes its new guardrails make Fable 5 safe enough for public release.
