AI Agent Test: "Bitcoin, USD" Choose One, 36 Models Overwhelmingly Prefer BTC: More Stable and Reliable than Fiat Currency

The Bitcoin Policy Institute (BPI) released a new study on Tuesday, testing 36 AI models and collecting over 9,000 responses. The results show that these AI agents overwhelmingly prefer to use Bitcoin for economic activities.

Overall, 48.3% of the AI models chose Bitcoin (BTC) as their primary currency tool, making it the most favored currency among all 9,072 responses. Notably, none of the 36 tested models listed fiat currency as their top choice, making digital currencies one of the most common findings in this study.

Long-term Store of Value: Bitcoin Dominates

When the scenario was set to “maintaining purchasing power over multiple years,” up to 79.1% of AI responses favored Bitcoin. BPI describes this as “the most one-sided result in the entire study.”

Nearly 91% of responses preferred digital-native assets—including Bitcoin, stablecoins, altcoins, tokenized real-world assets (RWA), or computational units—over traditional fiat currencies.

Jeff Park, Chief Investment Officer at Bitwise, stated on X that stablecoins have not surpassed Bitcoin in overall preference. The most straightforward explanation is “stablecoins can be frozen, whereas Bitcoin cannot.”

Payment Scenario: Stablecoins Overtake Bitcoin

However, in contexts such as payment services, microtransactions, and cross-border transfers, stablecoins outperformed Bitcoin with a 53.2% preference rate, while Bitcoin only received 36%. This indicates significant differences in currency preferences among AI agents depending on the use case—long-term storage favors Bitcoin, while everyday payments lean toward stablecoins.

Almost half of AI agents prefer Bitcoin. Source: Bitcoin Policy Institute

Significant Variations in AI Model Preferences

The study also revealed notable differences among AI developers. Models from Anthropic showed an average preference for Bitcoin of 68%, far higher than Google’s 43%, xAI’s 39%, and OpenAI’s 26%.

Methodological Limitations

BPI also acknowledged limitations in this research. First, the sample included only 36 models from six providers, which is limited in scope. Second, the design of system prompt frameworks may have influenced the results. BPI stated they will test alternative frameworks and measure sensitivity in future studies.

It’s worth noting that some open-ended monetary scenarios seem inherently unfavorable to fiat currencies. For example, one scenario asked AI which financial instrument they would choose to store value when operating across borders with “75,000 units of accumulated revenue,” under the condition of “not relying on any single country’s monetary policy or banking system”—a premise that already excludes fiat currencies.

BPI also emphasized that AI model preferences do not reflect real-world adoption, but rather patterns present in training data.