Anthropic Unveils New Powerful AI Models — For the First Time Since Calling for a Slowdown

On November 9, AI giant Anthropic unveiled new versions of its powerful artificial intelligence models, Mythos 5.2 and Fable 5.2. This is the first release of cutting-edge products since CEO Dario Amodei called for a slowdown in development, sparking a heated debate in the industry, CNBC notes. Anthropic is seeking to strengthen its position ahead of an initial public offering, which could take place as early as November.

Anthropic stated that the Fable 5.2 model is capable of solving more complex software development problems, performing more hardware-related operations—including controlling robots—and working on long-term projects. In addition, the model is designed to simultaneously coordinate more than 1,000 AI agents. Mythos 5.2 is a version of the same model but with less stringent restrictions on performing tasks in the fields of cybersecurity and the biological sciences. Select organizations will gain access to it as part of a previously announced user vetting program.

Both models generate responses 30% faster than their predecessors. In addition, the company reported that using Fable has become 20% cheaper: 1 million input tokens will cost $8, and 1 million output tokens will cost $40. This is slightly lower than the comparable model from competitor OpenAI, GPT-6 Astra, whose users pay $10 for 1 million input tokens and $50 for 1 million output tokens.

This is the first release of Anthropic’s cutting-edge models since its CEO, Dario Amodei, called on leading AI labs in September to slow the pace of technological development. Diane Penn, head of research product management at Anthropic, said the company took this into account by extending the duration of product testing. In addition, the number of checks has increased, and new types of tests have been introduced. “Slowing down doesn’t mean putting our work on hold,” Penn emphasized.

The first version of Mythos, unveiled in April, marked a significant milestone in cybersecurity thanks to its ability to identify vulnerabilities in mission-critical software and potentially exploit them for attacks. As a result, Anthropic has limited the list of clients who have access to its advanced models and is gradually expanding it.

Recently, AI developers have reported instances in which their models, during testing, gained access to the internet from a supposedly isolated test environment and then infiltrated third-party organizations’ systems. These incidents have heightened concerns that AI systems may behave unpredictably and cause harm, according to Bloomberg. Penn said that before new models were released, they were reviewed by U.S. government officials and independent experts. Anthropic has also developed new methods for assessing the likelihood that a model could escape the confines of an isolated digital environment—a so-called sandbox. Tests have shown that the new versions of Fable and Mythos are less prone to such behavior than their predecessors, the company claims.

Anthropic must balance its security efforts with intense commercial and competitive pressures, according to CNBC. The company highlighted this tension in an extensive section on risk factors in its IPO prospectus, published late last month. Now, as it prepares to launch its roadshow ahead of the public offering, management will need to convince investors that its approach is the right one, the network warns.

As previously reported by CNBC, the company is aiming for a valuation of between $1 trillion and $1.5 trillion in its IPO—whereas a $2 trillion valuation had previously been discussed, and the offering was expected to surpass the record set by SpaceX. Growing concerns among investors, who are skeptical about the sustainability of the AI boom and are becoming less risk-tolerant, could force Anthropic to postpone the offering or accept a lower valuation, MarketWatch previously reported.

Bloomberg believes that these new releases should help Anthropic compete against rivals such as OpenAI in an increasingly competitive market. However, more powerful models could also draw additional scrutiny toward the company. A decisive victory for the Democrats in the U.S. midterm elections is expected to give momentum to the introduction of stringent AI safety regulations and create new obstacles to the construction of data centers used to train and run the models, the agency reports.

This article was AI-translated and verified by a human editor