Anthropic has released to the public its top-tier "Mythos"-level artificial intelligence (AI) model, which it had delayed unveiling. It introduced safeguards to prevent misuse in sensitive areas such as cybersecurity, which had been a concern.
On the 9th (local time), Anthropic said it will launch the "Claude Fable 5," an AI model refined for safe public use at the Mythos level, and the security-specialized model "Claude Mythos 5." Anthropic said the etymology of "Fable" is the Latin fabula, with a meaning similar to "Mythos," which refers to myth. The two models are essentially the same, but Fable 5 is distinguished by safeguards applied to areas with a high risk of misuse.
When a query (prompt) comes in for the cybersecurity field that could be misused by malicious hackers, Fable 5 has the immediately lower-tier model, "Opus 4.8," which until recently was the top model, handle the response instead and informs the user of that fact.
Anthropic said, "We conservatively tuned these safeguards to release the model safely and quickly," adding, "Because of this, harmless requests can be blocked, but on average they are triggered in less than 5% of total sessions."
These safeguards also apply not only to cybersecurity but to biology and chemistry queries that could be misused for biological weapons, as well as to queries suspected of unauthorized "distillation," which extracts functions of competing AI models.
The unrestricted Mythos 5 will be selectively provided only to institutions vetted through the security consortium "Project Glasswing." Accordingly, in Korea, Samsung Electronics, SK hynix, SK Telecom, and the Korea Internet & Security Agency (KISA), which are known to participate in the project, appear likely to receive access.
Anthropic established a new data policy that retains data generated by the Fable and Mythos models for 30 days to use in defending against new attacks or identifying false positives.
Anthropic's new model posted even better performance benchmarks than the "Mythos preview" unveiled two months ago. In the "ExploitBench" assessment measuring cybersecurity-related capabilities, Mythos 5 scored 78%, outperforming OpenAI's GPT-5.5 at 34%, its own Opus 4.8 at around 40%, and even the Mythos preview at 69%.
In the "Humanity's Last Exam" (HLE), which gauges Ph.D.-level intelligence by field, it recorded 59% (without tool use), surpassing the Mythos preview at 56.8%, which first crossed the 50% threshold. In "Terminal-Bench 2.1," which measures coding ability in terminal environments, it also scored 88%, beating GPT-5.5 at 83.4%.
With Fable 5, safeguards are applied, so this level of performance is not directly perceptible. However, on "SWE-Bench Pro," which measures general coding ability, it scored 80.3%, higher than GPT-5.5 (58.6%) and Google Gemini 3.1 Pro (54.2%). In "GDPval-AA," which evaluates knowledge-work ability, it also scored 1,932 points, higher than GPT-5.5 (1,769 points) and Gemini 3.1 Pro (1,314 points).
Fable 5 is available starting today and will be provided to current paid subscribers without additional expense until the 22nd, but after that it will require a separate fee. Anthropic said it plans to include Fable 5 back in existing subscriptions once server capacity is sufficiently secured.