A powerful system was removed from corporate laboratories and handed over to ordinary users.
Tech companies have long been looking for a balance between open access to powerful AI models and controlling their possible consequences — and Anthropic solved this problem in its own way.
On June 9, the company released Claude Fable 5 is the first publicly available version of its flagship model Mythos. Prior to that, Mythos worked in a strictly limited circle: first, several partners, then in hundreds of organizations of 15 countries, mainly managers of critical infrastructure. Now access is open to everyone through APIs and corporate plans.
However, this openness has a rigid framework. In areas with high risk – cybersecurity, biology, chemistry – the model blocks the answers and switches to a less powerful Claude Opus 4.8. According to Anthropic, such cases are rare: at least 95% of sessions are held without forced switching.
Before the launch, the company conducted a large-scale security test: the internal program for finding vulnerabilities took more than 1000 hours and did not reveal universal ways to bypass the restrictions. External security checks teams have come to the same result.
Together with Fable 5, Anthropic has introduced a condition that can become a precedent for the entire industry. The company ordered the storage of traffic data for 30 days – even those customers who previously had agreements on the immediate deletion of information. Anthropic claims that the data will not go to training models, but will only be used to protect against new types of attacks and reduce false positives.
Access to the new model is paid: $10 per million input tokens and $50 per million weekends – twice as much as Opus 4.8. Until June 22, the Fable 5 is included in the subscriptions of Pro, Max, Team and Enterprise without extra charge, after which it will switch to the loan system. The company did not specify the terms of the model’s return to standard tariffs.
Tech companies have long been looking for a balance between open access to powerful AI models and controlling their possible consequences — and Anthropic solved this problem in its own way.
On June 9, the company released Claude Fable 5 is the first publicly available version of its flagship model Mythos. Prior to that, Mythos worked in a strictly limited circle: first, several partners, then in hundreds of organizations of 15 countries, mainly managers of critical infrastructure. Now access is open to everyone through APIs and corporate plans.
However, this openness has a rigid framework. In areas with high risk – cybersecurity, biology, chemistry – the model blocks the answers and switches to a less powerful Claude Opus 4.8. According to Anthropic, such cases are rare: at least 95% of sessions are held without forced switching.
Before the launch, the company conducted a large-scale security test: the internal program for finding vulnerabilities took more than 1000 hours and did not reveal universal ways to bypass the restrictions. External security checks teams have come to the same result.
Together with Fable 5, Anthropic has introduced a condition that can become a precedent for the entire industry. The company ordered the storage of traffic data for 30 days – even those customers who previously had agreements on the immediate deletion of information. Anthropic claims that the data will not go to training models, but will only be used to protect against new types of attacks and reduce false positives.
Access to the new model is paid: $10 per million input tokens and $50 per million weekends – twice as much as Opus 4.8. Until June 22, the Fable 5 is included in the subscriptions of Pro, Max, Team and Enterprise without extra charge, after which it will switch to the loan system. The company did not specify the terms of the model’s return to standard tariffs.
