Anthropic has just publicly released Fable 5, its most powerful Mythos-class AI model. Double the asking price

The company Anthropic, the creator of artificial intelligence (AI) models Claude, made the most powerful version of its technology available to the public on Tuesday, but still decided to limit access for certain sensitive areas, such as cyber security and – a novelty – the risks of biological and chemical attacks, writes AFP.

Called the Fable 5, this model is the first of the Mythos class – Anthropic's most advanced range, unveiled in April but restricted for security reasons – to be opened to the public.

In parallel, Anthropic proposes an unlimited version, Claude Mythos 5, for companies, organizations and government agencies that already have access to this family of models, presented as being able to detect and exploit security breaches with unprecedented speed and acuity.

In early April, Anthropic had simultaneously announced the existence of the Mythos class and its decision to allow access to it only to trusted partners in order to strengthen their cyber protection.

That news brought him accusations of “fear marketing” and at the same time caused a visible agitation among the leadership of countries and global institutions to assess the level of threat to essential infrastructures, such as banking, energy, etc.

Since then, several companies that had access to the Mythos range have confirmed the capabilities of these models, and the US Government, which is in conflict with Anthropic, ended up testing them.

The White House then instituted an optional device to evaluate the most powerful American AI models before they are commercialized. OpenAI, Anthropic's big rival, is about to unveil version 5.6 of its ChatGPT model, notes Agerpres.

Risks of biological and chemical attacks

Until now, Anthropic has justified restricting access to Mythos solely on the basis of cybersecurity risks.

With the launch of Fable 5, Anthropic therefore expands its vigilance and control over the risks of biological and chemical attacks.

The American company cited the example of the development of associated adenoviruses (AAV), potentially dangerous but also beneficial in gene therapy. Therefore, the unlimited version will be open to selected biology researchers, the company said.

According to its representatives, most requests for cybersecurity and biology are answered by an inferior model, Opus 4.8, which became public in late May.

Anthropic presents the Fable 5 as a secure version of the Mythos range, which without security measures could not be opened to the public. This framework relies on automated filters that analyze requests in real time and block sensitive ones.

Anthropic says it commissioned outside experts to try to counter those filters – an attack simulation procedure called “red teaming” – offering rewards to those who succeeded.

According to the American company, none of them – out of 1,000 hours of testing – managed “a universal bypass”, that is, a complete unlocking of the model. But “it's probably impossible to completely prevent universal circumventions,” Anthropic admits, with its goal being to make such attempts slow and expensive enough to detect before widespread deployment.

Also in the name of security, the American firm announced that it will keep all the data of these top models for 30 days.

Asking price for 1 million tokens

The issue of security, which Anthropic has made one of its main business arguments, has led the San Francisco-based startup to enter into an unprecedented dispute with the Trump administration over its refusal to lift restrictions on mass surveillance and lethal autonomous weapons.

After that, the Pentagon terminated contracts with the company, whose artificial intelligence tools were the only ones with a classified defense clearance.

The release of Fable 5 is accompanied by a high fee – up to $50 for a million “tokens”, the unit of measurement of requests and results produced – which is double the fee charged for Opus 4.8.

An intensive programming session can consume 1 million tokens in just a few hours or even less time.

Anthropic recently revised its pricing policy to account for the acceleration of AI agents, which often increase token consumption.

But the company has yet to prove its long-term profitability, hampered by the high cost of computing power required for AI tools.

Anthropic recently struck a deal with Elon Musk to use one of his data centers for $1.25 billion a month.

The launches of these products come at a time marked by financial effervescence around artificial intelligence.

Anthropic, like OpenAI, announced in June that it had filed for an IPO.

The two companies are trying to position themselves in the market just a few days before a record listing for SpaceX, Elon Musk's space giant, which also integrates its own artificial intelligence laboratory, called xAI.