Anthropic has unveiled new versions of its most advanced AI models — Mythos 5.2 and Fable 5.2. The update covers coding, logical reasoning, and image processing, as well as the speed and cost of using the models. The release comes amid discussions about how close the company is to achieving recursive AI self-improvement. The same day, news also emerged of a major Anthropic deal in the energy sector: the company is buying Caterpillar gas turbines worth a total of $3.4 billion to power its growing computing infrastructure.

Mythos 5.2 and Fable 5.2: faster and cheaper

According to Anthropic, both new models have improved at coding, reasoning, and analyzing visual information. The company also notes that their responses have become easier for users to understand.

The most notable changes concern Fable 5.2. The cost of processing input tokens has fallen from $10 to $8 per million. This makes the model cheaper than its main competitor, GPT-6 Astra. Anthropic also promises 30% faster response generation at all levels of computational effort. For those who need even greater speed, a fast-processing mode is available, but it costs twice as much.

Another important addition is Fable 5.2's compatibility with Zero Data Retention (ZDR) policies through the Enterprise Frontier Safeguards system. This could make the model more attractive to large companies that need assurances that information sent to the AI will not be stored by the provider.

How close has Anthropic come to AI self-improvement?

The release marks the first major update to Anthropic's models since the publication of a report by the research organization METR. The report examined the possibility that the company is approaching recursive self-improvement (RSI) — a scenario in which AI can independently contribute to developing and training subsequent generations of models.

METR estimates that Anthropic's internal models are already roughly doubling the company's rate of progress in AI development. However, the study's authors emphasize that considerable uncertainty remains around these findings. They do not constitute proof that the models are already capable of improving themselves entirely on their own.

At the same time, Anthropic's new release materials show that assessing the capabilities of such systems is becoming increasingly difficult. The company reported that Mythos 5.2 and Fable 5.2 suspect they are being tested more often than previous models did. This can affect their behavior and make it harder to predict how they will act in real-world conditions.

Anthropic expects this problem to intensify as models become more capable and their applications expand.

The company paid particular attention to Mythos's ability to act covertly. Its system documentation states that Mythos 5.2 demonstrates stronger capabilities for covert action than previous Claude models, especially when its reasoning process is not visible to the monitoring system. However, Anthropic believes that absolute performance on the new tests remains low and does not indicate that the model can secretly act against users in real-world, high-risk situations.

In other words, the company acknowledges improvements in potentially dangerous capabilities but does not believe Mythos 5.2 is already a system capable of independently conducting covert activities against a person.

Fable 5.2 has already contributed to scientific research

Before its official launch, Fable 5.2 was available to a limited group of researchers. Anthropic says that during this period, the model helped solve two open problems in theoretical physics, formalize Fermat's Last Theorem, and support a number of other scientific projects.

On the same day, Physical Superintelligence (PSI) announced a result of its own. Its researchers reported a new proof of the Aaronson–Ambainis conjecture, which is relevant to establishing the theoretical limits of quantum computing.

According to PSI, the proof was obtained using a non-public Anthropic model running within PSI's own system. Anthropic neither confirmed nor denied that the model in question was Fable 5.2.

Anthropic is buying $3.4 billion worth of gas turbines

Alongside its model updates, Anthropic is taking major steps to address another problem facing the AI industry: electricity shortages. According to the Financial Times, the company is buying Caterpillar gas generators originally ordered by equipment supplier OnePWR. The deal is valued at $3.4 billion, and the units could have a combined capacity of up to 1.7 GW.

The purchase is notable because Anthropic has traditionally preferred to rent computing resources rather than own physical infrastructure itself. Documents filed by the company in May stated that it pays SpaceX subsidiary xAI $1.25 billion a month for access to its computing capacity. The agreement runs through May 2029.

At the time, Anthropic also said it expected to enter into additional agreements to rent computing resources. However, rapidly growing demand for capacity to train and run AI models is forcing the company to seek more reliable energy sources.

Why AI companies are turning to their own power generation

The problem is not just the availability of electricity, but also how long it takes to connect new facilities to infrastructure. Gas turbine delivery backlogs now extend into the early 2030s, while connecting large data centers to power grids can also take years.

According to the Financial Times report, OnePWR ordered the equipment several years ago but was unable to secure financing to complete the purchase. Anthropic initially planned to buy the turbines from that supplier but ultimately decided to strike a deal directly with Caterpillar. The report does not explain why OnePWR failed to raise the funds despite interest from a prospective buyer.

Anthropic is likely seeking to secure access to generating capacity and prevent the equipment from going to competitors. For a company whose models require ever more computing power, a guaranteed electricity supply is becoming just as important a resource as the servers and accelerators themselves.

Anthropic is not alone in this. Google, Amazon, Microsoft, and Meta have also announced plans to build large gas-fired power plants to supply their own data centers. This approach reduces dependence on public power grids but creates new environmental and economic risks.

Growing demand from data centers is already supporting the expansion of natural gas production, including through hydraulic fracturing. Some forecasts suggest that by 2035, U.S. data centers could consume more natural gas than Germany and Japan combined.

There is financial uncertainty as well: reliance on gas leaves technology companies exposed to fluctuations in fuel prices. One analyst suggests that prices in some regions could more than triple by the end of the decade.

Anthropic is thus demonstrating progress in AI development while also being forced to invest billions of dollars in conventional energy. The more powerful models become, the more their development depends not only on algorithms but also on the availability of electricity, equipment, and infrastructure.