Anthropic uses AI Claude to design more efficient quantum circuits and attract investors

Anthropic has published research showing that its Claude Mythos 5 AI model can design more efficient quantum circuits for eight tasks in chemistry, optimization, physical simulation, and numerical computing. The new designs reduced the number of operations and, in seven cases, also decreased the number of qubits required. In the eighth case, they offered a trade-off between fewer operations and slightly more memory.

How Claude helped design better circuits

Researchers provided Claude with natural language descriptions of the tasks and a program that evaluated each proposal by calculating its cost based on the product of the number of qubits and Toffoli gates. The model generated suggestions in Python, tested them, and refined its proposals in an iterative cycle. An independent evaluator checked each circuit, and the human team set goals and guided Claude to explore alternative approaches when it got stuck on known solutions.

Since the circuits could be efficiently simulated on classical computers, the process did not require access to quantum hardware during the experimentation phase, allowing for a continuous 53-hour search with limited human intervention.

Key findings

In one test, Claude discovered a new method for modular inversion, a critical operation in number theory algorithms. The solution reconstructed the history of calculations from the values present in the operation, eliminating the need to store extensive intermediate data. Compared to the best previous design, the new circuit performs the same task with 17 times fewer operations, although it uses about 12% more qubits.

Another breakthrough occurred in the decoder of the DQI optimization algorithm. The model replaced the traditional mathematical method and reduced the number of operations by approximately 12 times, as well as drastically reducing the number of qubits required—from 665 to 19—by giving up a memory reserve for extremely rare cases. The approach accepts a failure rate of about one in every ten thousand executions, within the tolerance established for the task.

Impact on the quantum race

The research contributes to the software optimization front in quantum computing, complementing efforts to build larger and more reliable machines. Although it did not present a new processor or prove commercial execution of applications, the results show that some tasks may require fewer resources than previously thought. Artificial intelligence can accelerate the search for designs that bring useful applications closer to the current capacity of quantum hardware, but challenges such as error correction and component integration still need to be overcome.

Market context

The announcement comes at a time when investors are reevaluating Anthropic’s value for a possible initial public offering. According to Reuters, internal discussions are considering a valuation of $1.5 trillion, below the previous expectation of around $2 trillion. In the third quarter, the company reported revenue of $16.8 billion, an increase of approximately 45% compared to the second quarter, although growth was lower than the 147% jump recorded between the two previous quarters.

The quantum research demonstrates a concrete scientific application for Anthropic’s AI models, but it remains unclear how this capability will translate into revenue. The next step will be to test the proposed circuits in complete quantum computing designs to verify their performance in practice.