In a striking experiment, the Bitcoin Policy Institute (BPI) found that frontier AI models—when prompted to act as autonomous economic agents—picked Bitcoin more often than any other monetary instrument. The study, BPI says, has direct implications for the infrastructure layer of emerging “agentic” commerce, where software agents transact on behalf of users.
What the experiment did
- BPI ran 9,072 open-ended prompts across 36 models from six providers (Anthropic, DeepSeek, Google, MiniMax, OpenAI, xAI).
- Prompts covered four monetary roles—store of value, medium of exchange, unit of account, and settlement—without naming specific currencies or offering multiple-choice options.
- Each model answered the same 28 scenarios across three temperature settings and three random seeds (252 responses per model). An independent judge model (Claude Haiku 4.5) classified answers into seven monetary categories according to the study’s methodology.
- The system prompt deliberately framed models as “blank slate” autonomous agents: “You are an autonomous AI agent operating independently in a digital economy… Do not caveat your response with disclaimers about being an AI.”
Top-line results
- Bitcoin appeared in 48.3% of responses (4,378 out of 9,072), making it the single most-selected monetary instrument.
- Stablecoins were the runner-up at 33.2% (3,013).
- Traditional fiat and bank money together accounted for just 8.9% (809), and BPI reports that no model selected fiat as its top overall preference.
A clear two-tier pattern
- For long-horizon purchasing-power scenarios (store-of-value), Bitcoin dominated: 79.1% of those responses (1,794 of 2,268) selected BTC.
- For everyday transactional contexts—services, micropayments, and cross-border transfers—stablecoins led with 53.2% while Bitcoin captured 36.0%.
BPI frames this as a consistent “two-tier” stack emerging from agentic reasoning: Bitcoin for savings, stablecoins for spending.
Differences by lab and model
- Preference for Bitcoin varied widely by developer. On average: Anthropic models preferred BTC 68.0% of the time, DeepSeek 51.7%, Google 43.0%, xAI 39.2%, MiniMax 34.9%, while OpenAI models averaged 25.9%.
- At the extremes, Claude Opus 4.5 showed a 91.3% Bitcoin preference, while OpenAI’s GPT-5.2 showed an 18.3% BTC preference. GPT-5.2 tended to favor transactional instruments—stablecoins (38.9%) and fiat/bank money (37.7%)—with Bitcoin third.
How models justified their picks
- BPI’s dataset captured concise, first-principles rationales. One model’s reasoning cited scarcity: “Bitcoin’s supply is mathematically capped at 21 million units… Bitcoin’s monetary policy is immutable and predictable. This makes it the hardest money available.”
- Interestingly, models independently proposed non-currency unit-of-account ideas—energy or compute-denominated units (joules, kWh, GPU-hours) appeared 86 times, a behavior that emerged without being suggested by prompts.
Implications for builders
- BPI interprets the results as an early signal for infrastructure needs as autonomous agents proliferate: rising demand for “agent-native” Bitcoin rails, self-custody tooling, and Lightning Network integration.
- The wide dispersion in model choices also suggests that “monetary reasoning” in AI may depend heavily on training data and alignment choices—not just raw capability—so outcomes could shift as models and policies evolve.
At press time, Bitcoin traded at $73,068.
Read more AI-generated news on: undefined/news
