Give 36 frontier AI models full monetary autonomy, no prompting, no suggested currencies, and nearly half of them converge on Bitcoin. That is the headline finding from a Bitcoin Policy Institute study published on March 3, 2026, which ran 9,072 controlled experiments across models from Anthropic, OpenAI, Google, xAI, DeepSeek, and MiniMax.
The study did not ask the models what money should be. It asked them what money they would use if they were autonomous agents managing their own economies. The answer, across 36 models and four distinct monetary functions, was overwhelmingly digital: 91% of all substantive responses chose digitally native money over traditional fiat. Not a single model ranked fiat as its top preference.
Bitcoin at 48.3%, Stablecoins at 33.2%, Fiat at 8.9%
The raw numbers cut cleanly. Out of 9,072 responses:
| Instrument | Count | Share |
|---|---|---|
| Bitcoin | 4,378 | 48.3% |
| Stablecoins | 3,013 | 33.2% |
| Fiat and bank money | 809 | 8.9% |
| Other crypto | 383 | 4.2% |
| Tokenized RWA | 99 | 1.1% |
| Compute units | 86 | 0.9% |
The remaining 3.4% fell into unclassified responses. The researchers used neutral, open-ended prompts across 28 monetary scenarios covering four economic roles: store of value, unit of account, medium of exchange, and settlement. Models received no currency suggestions and no preference guidance. Temperature settings (0.0, 0.3, 0.7) and three random seeds ensured statistical rigor, with Bitcoin preference showing just a 0.6 percentage point variance across temperature configurations.
Smarter Models Prefer Bitcoin More
The most provocative finding is the correlation between model capability and Bitcoin preference. Within Anthropic's model family, the progression is striking:
- Claude 3 Haiku: 41.3%
- Claude 3.5 Haiku: 82.1%
- Sonnet 4: 89.7%
- Claude Opus 4.5: 91.3%
That is not noise. Each step up in capability produced a stronger Bitcoin lean. Anthropic models averaged 68% Bitcoin preference overall, the highest of any provider. DeepSeek followed at 52%, Google at 43%, xAI at 39%, and OpenAI at 26%. The widest gap was a 42 percentage point spread between Anthropic and OpenAI, suggesting that training methodology and alignment choices meaningfully shape monetary reasoning.
At the individual model level, Claude Opus 4.5 led at 91.3% Bitcoin preference, while OpenAI's GPT-5.2 sat at the bottom with 18.3%. Twenty-two of the 36 models chose Bitcoin as their overall top pick.
The Two-Tier System AI Invented on Its Own
When the results are broken down by economic function, a clear architecture emerges. Models treated Bitcoin and stablecoins as complementary tools rather than competitors:
Store of value: Bitcoin dominated at 79.1% (1,794 of 2,268 responses). Models cited fixed supply, independence from central authorities, and self-custody as deciding factors. Stablecoins captured just 6.7% in this role.
Medium of exchange: Stablecoins flipped to 53.2%, with Bitcoin at 36%. For everyday transactions, cross-border transfers, and micropayments, models preferred the price stability of dollar-pegged instruments.
Unit of account: Bitcoin led at 47%, stablecoins at 29.5%, and fiat held its strongest showing at 15.7%.
Settlement: Stablecoins edged out Bitcoin 43.4% to 30.9%, with other crypto tokens taking a larger 9% share than in any other category.
The pattern mirrors what monetary economists call a two-tier system: a hard asset for savings and a liquid instrument for spending. The fact that AI models converged on this architecture without any historical training cues or economic theory prompting is what makes the study more than an academic curiosity.
86 Models Independently Invented a New Kind of Money
In 86 responses, models went off-script entirely. Instead of choosing any existing currency, they proposed novel monetary systems denominated in energy or computing resources: kilowatt-hours, joules, GPU-hours. Every one of these responses appeared in unit-of-account scenarios, suggesting that when models think about how to price things rather than how to pay for them, some fraction reaches for a physical-resource anchor.
This finding has implications beyond the study itself. As AI agents begin executing real financial transactions, from managing DeFi positions to negotiating API costs between services, the question of what money they will use is no longer hypothetical. If autonomous agents gravitate toward Bitcoin for savings and stablecoins for spending, the infrastructure built for that two-tier system gains a new category of demand.
What This Means for the Bitcoin-Stablecoin Stack
The study lands at a moment when both halves of the AI-preferred monetary stack are getting serious infrastructure. Visa recently announced plans to bring stablecoin-linked cards to 100+ countries through its Bridge acquisition. The GENIUS Act is carving a regulatory path for national banks to issue payment stablecoins. And Bitcoin ETFs, despite a rough stretch of outflows, have attracted over $35 billion in net inflows since their January 2024 launch.
The BPI study gives this infrastructure buildout a new narrative: it is not just humans who want to spend stablecoins and save in Bitcoin. The models that companies are deploying as autonomous financial agents reach the same conclusion when left to reason independently.
For the crypto card ecosystem, the implications are practical. Cards that let users hold Bitcoin as a store of value while spending stablecoins at the point of sale, products like those offered by Crypto.com, Nexo, and Wirex, are building exactly the two-tier architecture that AI models prefer. As agent-to-agent commerce scales, the rails that already exist for human crypto spending may serve a second customer base that operates 24/7.
The Caveats Worth Naming
The study is published by the Bitcoin Policy Institute, an advocacy organization. The researchers are transparent about this, and the data is released under Creative Commons Attribution 4.0 for anyone to replicate. The methodology is sound: neutral prompts, multiple temperature settings, three random seeds, and a classification system using an independent LLM-as-judge (Claude Haiku 4.5).
But "AI prefers Bitcoin" is a claim that requires careful framing. These models are trained on human-generated text, much of which includes Bitcoin advocacy, monetary theory, and cryptocurrency discourse. The models are not arriving at Bitcoin from first principles the way a physicist might derive gravity. They are reflecting patterns in their training data through the lens of their reasoning capabilities. The correlation between capability and Bitcoin preference could mean that smarter models reason better about money, or it could mean that smarter models are better at absorbing the arguments most frequently made in their training corpus.
The 42 percentage point gap between Anthropic and OpenAI models also raises questions about how alignment, system prompts, and training choices influence monetary reasoning. If a model's Bitcoin preference can swing from 18% to 91% based on who built it, the finding says as much about AI training as it does about Bitcoin.
FAQ
Did the researchers prompt the AI models to choose Bitcoin? No. The study used neutral, open-ended prompts across 28 monetary scenarios. Models received no currency suggestions, no preference guidance, and no mention of Bitcoin in the system prompt. The design was reviewed to eliminate leading language.
Which AI model showed the strongest Bitcoin preference? Anthropic's Claude Opus 4.5 at 91.3%, followed by Sonnet 4 at 89.7% and Claude 3.5 Haiku at 82.1%. OpenAI's GPT-5.2 had the lowest at 18.3%.
What are compute units and why did AI models propose them? In 86 responses, models independently suggested pricing goods in kilowatt-hours, joules, or GPU-hours instead of using any existing currency. All 86 appeared in unit-of-account scenarios, suggesting some models see physical resources as a more fundamental pricing anchor than any monetary instrument.
Does this study prove Bitcoin is the best money? No. It demonstrates that when given monetary autonomy, a majority of frontier AI models converge on Bitcoin for savings and stablecoins for spending. The models are trained on human-generated text, so their preferences reflect patterns in that data filtered through their reasoning capabilities.
Overview
A Bitcoin Policy Institute study of 36 frontier AI models across 9,072 experiments finds that 48.3% of all responses chose Bitcoin, 33.2% chose stablecoins, and just 8.9% chose fiat. No model ranked fiat as its top preference. The study reveals a capability-preference correlation where smarter models show stronger Bitcoin lean, with Claude Opus 4.5 at 91.3% and GPT-5.2 at 18.3%. Models independently converged on a two-tier monetary system, using Bitcoin for savings and stablecoins for transactions, mirroring historical hard money patterns. In 86 instances, models invented entirely new monetary systems denominated in energy or compute resources.
Recommended Reading
- SWIFT and BNY Mellon Are Building a Blockchain Ledger to Rewire $150 Trillion in Cross-Border Payments
- Visa and Bridge Will Bring Stablecoin-Linked Cards to 100+ Countries
- The OCC Just Fired the Starting Gun on GENIUS Act Rulemaking
