What is Claude Fable 5? Anthropic’s Mythos-level AI comes with built-in safety brakes

Anthropic on Tuesday launched Claude Fable 5, its most powerful generally available artificial intelligence model to date, claiming it outperforms all of its previous systems across software development, scientific research, data analysis and other complex tasks.

According to the company, Fable 5 can handle long and complex assignments more effectively than previous generations, demonstrating stronger abilities in software engineering, business analysis, vision-based tasks, memory and scientific research.

The launch builds on Anthropic's Project Glasswing initiative, announced in April, which was created to help governments, technology companies and critical infrastructure providers prepare for a future in which AI systems can identify software vulnerabilities more effectively than most human experts.

Anthropic had warned that AI models had reached a level where they could outperform all but the most skilled cybersecurity professionals in discovering and exploiting software flaws.

The initiative initially brought together around 50 organisations, including major technology firms Amazon Web Services, Apple, Google, Microsoft and NVIDIA, as well as financial giant JPMorgan Chase and cybersecurity companies CrowdStrike and Palo Alto Networks.

Anthropic said internal testing of its earlier Claude Mythos Preview model had already uncovered thousands of high-severity software vulnerabilities, including weaknesses affecting major operating systems and web browsers.

According to the company, partners participating in Project Glasswing had identified more than 10,000 serious security vulnerabilities by late May that could potentially have been exploited by hackers.

The programme was originally limited to a small group of mostly US-based organisations while Anthropic evaluated the risks associated with the technology.

Last week, the company expanded access to approximately 150 organisations across more than 15 countries.

The new participants include operators of power grids, water systems, hospitals, telecommunications networks and hardware manufacturers.

Anthropic said many of these organisations provide software and services that are relied upon by governments and millions of businesses and consumers worldwide.

A successful cyberattack against some of these systems could potentially affect more than 100 million people and have significant implications for national and global security, the company said.

Safeguards to curb misuse

Anthropic said the model's advanced capabilities also create new risks, particularly in cybersecurity and biological research, prompting the company to introduce a new layer of safeguards.

Fable 5 will automatically redirect certain sensitive queries to an older but safer model, Claude Opus 4.8.

The safeguards are designed to prevent users from exploiting the model's capabilities for cyberattacks, dangerous biological research or attempts to copy and replicate the model's capabilities.

The company said these safeguards are intentionally conservative and may occasionally block harmless requests.

However, it estimated that fewer than 5 per cent of user sessions would be affected by the fallback mechanism.

"More than 95 per cent of Fable sessions involve no fallback at all," Anthropic said, adding that users would be informed whenever their request is being handled by Opus 4.8 instead of Fable 5.

“Opus 4.8 is a highly capable model in its own right: a response that falls back to Opus is a far better experience than an outright refusal from Fable,” it said.

How’s Fable 5 more capable?

In one early test, payments company Stripe reportedly used Fable 5 to complete a migration across a 50-million-line software codebase in a single day, a task that would otherwise have taken a team of engineers more than two months.

The company also highlighted gains in business and financial analysis, saying Fable 5 achieved leading scores on benchmarks that measure document analysis, reasoning and decision-making capabilities.

In visual tasks, Anthropic said the model can accurately extract information from scientific charts and figures and can even recreate software code from screenshots of applications.

The company also pointed to improvements in long-term memory and context handling, saying Fable 5 can remain focussed across millions of tokens of information and improve its performance using notes created during earlier stages of a task.

Claude Mythos 5 reserved for trusted users

The company also unveiled Claude Mythos 5, a version of the same underlying model with fewer restrictions, which will remain available only to a limited group of trusted cybersecurity professionals and critical infrastructure operators.

Anthropic said Claude Mythos 5 possesses the most advanced cybersecurity capabilities of any AI model currently available.

The company said Mythos 5 has also shown promising results in drug discovery and biological research.

According to Anthropic, internal researchers found that the model accelerated parts of the drug-design process by around ten times.

In several tests, the model independently selected biological targets, operated specialised research tools and generated candidate drug designs with performance comparable to skilled human researchers.

The company further claimed that Mythos 5 can generate original scientific hypotheses and conduct extended research projects with limited human supervision.

In one experiment, the model reportedly spent more than a week conducting genomics research, analysing data from millions of cells across 138 animal species and creating a machine-learning system that outperformed a recently published scientific model despite being significantly smaller.

Anthropic, however, acknowledged that such advanced capabilities could potentially be misused.

To address these concerns, the company said it had developed specialised AI-powered classifiers that monitor requests for signs of cyberattacks, dangerous biological research or efforts to extract the model's capabilities for training rival systems.

Anthropic said extensive internal and external testing found its cybersecurity safeguards to be highly resistant to so-called "jailbreaks", attempts to bypass AI safety restrictions.

The company also announced a new data retention policy for its most advanced models.

Business customers using Fable 5, Mythos 5 and future Mythos-class systems will be subject to a mandatory 30-day data retention period.

Anthropic said the information would not be used to train future models and would be retained solely for safety and security purposes before being deleted.

Pricing, availability and rollout plans

The company priced both Fable 5 and Mythos 5 at USD 10 per million input tokens and USD 50 per million output tokens, less than half the price of its earlier Mythos Preview model.

Claude Fable 5 is available immediately worldwide through Anthropic's platform and application programming interface (API).

Access to Mythos 5, however, will remain restricted to participants in Anthropic's Project Glasswing cybersecurity initiative and, in the coming weeks, a select group of life-sciences researchers under a new trusted-access programme.

Anthropic said it plans to gradually expand access to Mythos 5 in consultation with the US government while continuing to strengthen its safety mechanisms as AI systems become more powerful.