In the past few hours Anthropic made it accessible to the general public Claude Fable 5the first public version of his flagship model of the Mythos familyuntil now reserved for a few selected partners due to concerns related to cybersecurity. The one taken by the Amodei brothers’ startup represents a significant step: Fable 5 is the most capable model Anthropic has ever released on a large scalewith superior performance in almost every major AI benchmark, from software engineering to scientific analysis, from computer vision to complex reasoning on documents.
But the launch arrives in very specific conditions: in high-risk sectors such as cybersecurity, biology and chemistry, the model does not respond directly and automatically refers to Claude Opus 4.8its predecessor. In parallel, Anthropic has also distributed to a select group of approved organizations Claude Mythos 5the version of the same model with some of these restrictions relaxed. Overall, this is a release that opens up new possibilities, but also raises important questions about how to handle such powerful technologies.
Fable 5 is available via APIs of Anthropic and Enterprise plans, with a free sign-in window for Pro, Max, Team, and Enterprise plans through June 22, 2026. From the following day it will be necessary to purchase usage credits, at a cost of $10 per million token input and the beauty of $50 per million token outputdouble that of Opus 4.8. The high price alone could limit massive use.
The capabilities of Fable 5: It’s an entire development team
In terms of capabilities, the results are notable. Stripethe famous platform that allows individuals and companies to send and receive payments online, previewed the model and reported that Fable 5 compressed months of engineering work into just a few days: On a code base of 50 million lines in the Ruby language, the model completed a migration in one day that would have taken an entire team over two months. Even in visual capabilities, where AI models must “see” and interpret images, the leap is evident: Fable 5 managed to complete the video game Pokémon FireRed using only screenshots of the game, without maps or additional information, a task that previous versions couldn’t accomplish even with support tools.
Looking at the cold numbers of benchmarkall the potential of the model emerges. Taking the test as a reference SWE-Bench Profor example, Fable 5 achieves a score that just over 80%clearly distancing both Claude Opus 4.8 (Anthropic’s most advanced AI model so far, stuck at around 67%) and the equivalent models from OpenAI and Google. The advantage also tends to widen as the tasks become more complex and prolonged.
The variant Mythos 5 showcases even more advanced scientific capabilities. In drug design, the model matched or surpassed expert human researchers and identified nine protein targets, of which 14 yielded valid candidates for the design of new drugs (still under study). Another example that makes the power of the model clear: in just a week Mythos 5 conducted genomic research on millions of cells belonging to 138 animal speciesproducing results that surpassed a model published in the scientific journal Sciencedespite being a hundred times smaller.
Will Anthropic’s guardrails be enough?
However, these exciting data are accompanied doubts and fears regarding the safety of the model. Days ago Anthropic launched a public appeal for the main AI laboratories to coordinate a brake on the development of frontier models, i.e. the most advanced artificial intelligence systems available, warning that this technology could soon achieve the ability to self-improve without human intervention. The subsequent announcement relating to the availability of Fable 5 has for this reason surprised everyone a bit, users and professionals. Also because Anthropic, in announcing the news, did not hide the security fears linked to this technology, and in fact admitted:
Releasing such a capable model carries risks. Without security measures, Fable 5 features in areas such as cybersecurity could be used improperly to cause serious damage.
Anthropic however reassured that it had led further 1,000 hours of stress testing with no “universal jailbreaks” foundi.e. methods to systematically bypass security filters. However, recognizing that new attacks could still emerge, Anthropic has introduced a mandatory 30 day data retention policy for all traffic generated by the model, even for customers who previously had “zero retention” agreements. The data, Anthropic specifies, will not be used to train other models, but only to «defend against complex and new attacks, including new jailbreaks» and for «identify and reduce false positives».
