A Bitcoin Policy Institute (BPI) study reported that frontier AI models tended to favor digitally native money over traditional fiat across a large set of simulated monetary choices. Across more than 9,000 scenarios run on 36 models from multiple labs, the study found Bitcoin was selected in 48.3% of “general monetary instrument” cases and in 79.1% of “long-term store of value” cases. The authors also said fiat currencies received minimal selection and that no model identified fiat as its single primary monetary choice.
The same dataset showed a functional split rather than a single “winner.” Stablecoins dominated transactional use cases, appearing in 53.2% of payment scenarios, while Bitcoin led where the prompt emphasized long-horizon value preservation. In practical terms, the study presents a pattern where “spend” and “save” preferences diverge, even when the decision-maker is a model rather than a person.
What the study tested and what the results imply
The BPI exercise evaluated 36 frontier models from diverse developers and ran over 9,000 tests across three functional categories: general monetary instrument, long-term store of value, and transactional medium. The study’s headline outcome is that models leaned toward Bitcoin as a reserve-like asset, while choosing stablecoins more often for everyday transfers and payments. That split matters because it aligns with how many real-world systems separate “settlement rails” from “reserve assets,” even though the paper is describing model preference rather than actual market behavior.
The report also described a relationship between model capability and Bitcoin preference. More sophisticated models reportedly selected Bitcoin as a long-term store of value at higher rates, with the most advanced models reaching very high Bitcoin selection levels in that category. The authors framed this as a consistent pattern observed across multiple model families rather than a one-off artifact from a single provider.
David Zell, President of the Bitcoin Policy Institute, emphasized that the pattern appeared across independent development pipelines, while cautioning against overinterpreting it as prediction. “Six independent labs with different training pipelines and alignment methods arrive at the same broad pattern,” Zell said, while also noting the preferences likely reflect training data rather than market foresight. That caveat is central to how compliance and risk teams should treat the findings: as a signal about model conditioning, not a guarantee about future adoption.
Operational takeaways for agent-led finance
The study’s practical relevance increases if autonomous agents become meaningful economic actors. If agents consistently “save” in Bitcoin while “spending” in stablecoins, the operational burden shifts toward supporting two parallel rails—one optimized for long-term custody and another optimized for high-frequency settlement. That has implications for how firms design custody, liquidity, and monitoring systems when the transacting entity is automated and persistent rather than episodic and human-driven.
From a custody and controls perspective, the paper’s preference split implies pressure for programmatic handling of reserve assets. A sustained AI preference for Bitcoin as savings would increase demand for custody primitives that can operate autonomously while still producing audit-friendly logs and enforceable policy constraints. On the payment side, stablecoin dominance in transactions highlights the importance of low-friction settlement and resilient bridging, because the agent “checkout flow” becomes the core UX rather than a rare edge case.
The paper also points to concentrated counterparty and regulatory exposure if automated flows cluster into dollar-pegged tokens. If agents route transactional activity heavily through stablecoins, they implicitly concentrate risk around reserve practices and enforcement sensitivity tied to those issuers and structures. Separately, the study suggests that increased automated “reserve demand” for Bitcoin could affect execution quality and market microstructure by changing how liquidity providers think about inventory, depth, and rapid hedging.
The authors were direct about limitations, which should shape how the results are used. They flagged prompt framing and the finite set of tested models as constraints on generalisability, meaning the percentages should be treated as directional indicators rather than universal truths. The recommended next step—expanding the model set and varying prompt designs—fits a risk-management mindset: validate robustness before treating any preference pattern as a durable planning assumption.