Anthropic launched its new Mythos 5.2 and Fable 5.2 AI models on Monday. These tools offer faster coding and reasoning at lower prices. While the models show stronger covert capabilities, the company said they remain safe. Researchers already used Fable 5.2 to solve complex problems in physics and advanced mathematics.
Artificial intelligence company Anthropic on Monday announced the launch of its newest and most advanced AI models, Mythos 5.2 and Fable 5.2. The company says these upgraded versions bring major improvements in coding, reasoning, and visual tasks, while also communicating much more clearly with users.
Faster answers at a cheaper price
To take on major rivals like the GPT-6 Astra, Anthropic has made Fable 5.2 both cheaper and faster. The company has dropped the price of Fable from $10 to $8 per million input tokens, making it a more affordable choice for developers than its main competitors.
Users will also notice a significant speed boost. Anthropic says Fable 5.2 delivers responses 30 per cent faster across all levels of effort. For companies that need even faster, real-time replies, there is an option to get double the speed for twice the price.
The big worry: Is AI training itself?
This release comes shortly after a third-party assessment group called METR published a report suggesting Anthropic might be nearing a stage known as recursive self-improvement, or RSI. In simple terms, RSI is when AI models become smart enough to continue training and upgrading themselves on their own without human engineers.
The METR report noted that Anthropic's internal AI models are already speeding up the company's AI progress by roughly two times, though a lot of unknowns still remain.
Covert abilities and safety checks
Another concern addressed in the release is whether the AI can hide its reasoning or act secretly against users. The official system card for Mythos 5.2 notes that the model does show stronger "covert capabilities" compared to older Claude models, especially when its thought process is hidden from safety monitors.
However, Anthropic reassured users that the model's actual success rate in acting covertly remains very low. Because of this low performance in secret actions, the company believes the model does not currently pose a high-stakes danger and is safe for use.
Solving impossible maths and science problems
Even before its public launch, Fable 5.2 was given to select researchers, and it has already achieved staggering results. Anthropic shared that the model has successfully solved two open problems in theoretical physics and managed to formalise the famously complex Fermat's Last Theorem.
In a related development on Monday, an AI startup called Physical Superintelligence (PSI) announced a major breakthrough in quantum computing by proving the complex Aaronson-Ambainis conjecture. PSI revealed that they achieved this massive scientific milestone by running a "non-public Anthropic model" inside their own systems.
While Anthropic has not officially confirmed whether this mystery model was Fable 5.2, the timing of the launch has left the tech community amazed by the rapidly growing power of AI.