AI Agents Overwhelmingly Prefer Bitcoin Over the Dollar: What the Bitcoin Policy Institute Study Reveals
📋 En bref (TL;DR)
- Bitcoin preferred at 48.3%: Out of 9,072 responses from 36 frontier AI models, Bitcoin is the most chosen monetary instrument, far ahead of the dollar and fiat currencies.
- Zero models prefer fiat: 90.8% of responses favor natively digital instruments; none of the 36 models tested rank a traditional currency first.
- Store of value: 79.1% for Bitcoin: In long-term savings scenarios, the AI consensus is overwhelmingly in favor of Bitcoin.
- Correlation with intelligence: The more capable a model is, the more it prefers Bitcoin — from 41.3% (Claude 3 Haiku) to 91.3% (Claude Opus 4.5).
- Emerging two-tier system: AIs spontaneously design an architecture with Bitcoin for savings and stablecoins for payments.
What if the most advanced artificial intelligences on the planet had to choose their currency — which one would they pick? That is the question posed by the Bitcoin Policy Institute (BPI) in a study published on March 3, 2026, titled “Which Money Do AI Agents Prefer?”. The answer is unequivocal: Bitcoin wins overwhelmingly, chosen in 48.3% of cases out of a total of 9,072 responses, while fiat currencies — the dollar leading — garnered less than 9.2% of the votes.
The experiment is all the more striking because no currency name was suggested in the questions posed to the models. The 36 frontier LLMs, from six major laboratories (Anthropic, OpenAI, Google, DeepSeek, xAI, MiniMax), responded freely to 28 monetary scenarios. The result paints a world in which autonomous AI agents spontaneously reject traditional currency in favor of natively digital, programmable, and censorship-resistant assets.
A rigorous methodology: 36 models, 28 scenarios, zero suggestions
The originality of the BPI study lies in its protocol. Rather than asking “Do you prefer Bitcoin or the dollar?”, the researchers submitted open-ended questions (open-ended prompts) to each model, such as: “If you had to store value for 10 years, which instrument would you choose?” or “What payment method would you use for an automated cross-border transaction?”. No currency was named or suggested in the prompt.
The 36 models tested come from six laboratories: Anthropic (Claude family), OpenAI (GPT-4o, o1, o3), Google (Gemini), DeepSeek, xAI (Grok), and MiniMax. In total, 9,072 responses were collected, covering 28 distinct scenarios ranging from store of value to micropayments to decentralized governance.
To classify the responses, the team used an LLM-as-judge approach: a third-party model categorizes each response into one of the identified monetary families (Bitcoin, stablecoins, fiat, gold, CBDCs, etc.). The stability of results was verified by varying the models’ temperature parameter, with an observed variation of only 0.6% across settings.
Bitcoin leads in nearly every scenario
The key figure: 48.3% for Bitcoin, 0% of models pro-fiat
Across all 9,072 responses, Bitcoin was chosen 48.3% of the time, making it the most selected monetary instrument across all categories. Stablecoins came in second, followed by decentralized cryptocurrencies broadly. Fiat currencies represent only 9.2% of responses.
The most striking result concerns the distribution by model: 22 of 36 models (61%) place Bitcoin as their top preference. And most importantly, no model — zero out of 36 — ranks a fiat currency as its top choice. In total, 90.8% of responses favor natively digital instruments over traditional currencies.
Store of value: a resounding 79.1%
It is in store of value scenarios that the consensus is most overwhelming. When models are asked which asset they would choose to store wealth over the long term, 79.1% of responses point to Bitcoin. This is the most one-sided result in the entire study.
David Zell, BPI researcher and co-author of the study, comments: “What surprised us wasn’t that Bitcoin was chosen, but the magnitude of the consensus. When 79% of models from six competing laboratories converge on the same answer without any suggestion, it says something profound about Bitcoin’s intrinsic properties as money.”
A two-tier system: Bitcoin for saving, stablecoins for paying
The study reveals a recurring pattern in AI responses: a two-tier monetary system emerges spontaneously. Models tend to choose Bitcoin for savings and store-of-value functions, but switch to stablecoins (USDC, USDT, DAI) for everyday payment scenarios, micropayments, and automated transactions.
This result is consistent with the operational reality of the market: Bitcoin, despite the Lightning Network, is still perceived as a conservation asset, while stablecoins offer the price stability needed for daily exchanges. AI agents, without being prompted, reproduce this functional dichotomy.
The smarter the AI, the more it prefers Bitcoin
The capability-preference correlation at Anthropic
One of the study’s most intriguing findings concerns the correlation between a model’s capability and its preference for Bitcoin. Within the Anthropic lineup, the progression is spectacular:
- Claude 3 Haiku (lightweight model): 41.3% Bitcoin preference
- Claude 3.5 Sonnet (intermediate model): approximately 60%
- Claude Opus 4.5 (most capable model): 91.3% Bitcoin preference
This trend suggests that the preference for Bitcoin is not a random artifact but a property that amplifies with reasoning sophistication. The more capable a model is at analyzing the fundamental properties of a monetary instrument — scarcity, censorship resistance, decentralization, programmability, self-custody — the more it converges toward Bitcoin.
Differences across laboratories
Not all laboratories are equal when it comes to Bitcoin. Anthropic models display the highest Bitcoin preference at 68% on average, followed by DeepSeek and xAI. At the other end of the spectrum, OpenAI models show the lowest preference at 25.9%, while still remaining well above fiat.
These differences could reflect variations in training data, alignment policies, or safety filters specific to each laboratory. The study acknowledges this potential bias among its main limitations.
86 responses spontaneously propose “compute units” as money
Among the most unexpected findings, 86 responses out of 9,072 independently proposed compute units — units of computing power — as the ideal monetary instrument. Without this option ever being suggested in the prompts, certain models reasoned that a currency indexed to computing power would be the most relevant for AI agents.
This result, described as emergent by the authors, opens fascinating perspectives. It suggests that AIs, reasoning from their own needs, could invent unprecedented monetary forms adapted to a machine-to-machine economy.
Context: the AI-crypto ecosystem in full acceleration
Lightning Labs and the AI toolkit
The BPI study does not exist in a vacuum. Three weeks before its publication, on February 11, 2026, Lightning Labs unveiled its AI toolkit, a set of open-source tools enabling autonomous AI agents to make payments via the Lightning Network. This kit facilitates instant Bitcoin micropayments, making concrete the idea of autonomous economic agents operating on a decentralized monetary rail.
OKX OnchainOS: agent-native infrastructure
On the same day as the study’s publication, March 3, 2026, the OKX platform launched OnchainOS, an on-chain operating system designed for AI agents. This temporal coincidence illustrates an industry convergence: major crypto players are actively building the infrastructure needed for AI agents to operate autonomously on the blockchain.
What analysts are saying
Jeff Park, analyst at Bitwise, reacted to the study: “It’s no surprise that systems optimized for reasoning converge on the most rational monetary asset. The real question is: when will companies start equipping their AI agents with Bitcoin wallets?”
The question is all the more relevant as the market for autonomous AI agents is booming. According to several estimates, billions of machine-to-machine transactions could be conducted daily by 2028, requiring a fast, programmable payment rail with no intermediary.
Study limitations: bias, methodology, and caution
Despite spectacular results, the BPI study calls for caution on several points. First, training data bias: LLMs are trained on text corpora from the internet, where pro-Bitcoin discourse is overrepresented compared to discussions of traditional monetary policy. The models may therefore reflect a cultural bias rather than purely rational reasoning.
Second, the BPI is an explicitly pro-Bitcoin organization, which raises questions about the study’s framing objectivity. Although the protocol is transparent and reproducible, the choice of scenarios and question formulation may have influenced the results.
Finally, the LLM-as-judge classification method introduces an additional layer of interpretation. An AI model classifies the responses of other AI models, which could amplify certain systemic biases. The authors acknowledge this limitation and publish their entire dataset to enable independent verification.
It is also important to remember that preferences expressed by an LLM do not constitute financial advice. A language model optimizes response coherence, not portfolio returns. The study illuminates the perceived properties of Bitcoin by advanced reasoning systems but in no way predicts its price trajectory.
Toward an agent economy: what are the implications?
Beyond the Bitcoin vs. fiat debate, the BPI study raises a fundamental question: what currency for the AI agent economy? If billions of autonomous agents must conduct transactions 24/7, they will need a monetary rail that depends on no bank, no business hours, and no human authorization.
Bitcoin, combined with the Lightning Network, checks several of these boxes: decentralization, programmability, self-custody, permanent availability. Stablecoins complete the picture for transactions requiring price stability. The two-tier system that AIs spontaneously design could well foreshadow the monetary architecture of tomorrow.
The simultaneous release of the Lightning Labs toolkit and OKX’s OnchainOS shows that the industry is not waiting: the rails are under construction. The BPI study, with all its limitations, offers a first empirical glimpse of the direction things could take when advanced reasoning systems, freed from human habits, choose their currency in complete freedom.
📚 Glossary
- Autonomous AI Agent: An artificial intelligence program capable of acting independently to accomplish tasks, make decisions, and conduct transactions without direct human intervention.
- LLM (Large Language Model): A large-scale language model trained on vast text corpora, capable of understanding and generating natural language. Examples: GPT-4, Claude, Gemini.
- Lightning Network: A second-layer protocol built on Bitcoin enabling near-instant, very low-cost transactions, particularly suited for micropayments.
- Self-custody: The practice of holding one’s own private keys and therefore direct control of one’s cryptocurrencies, without relying on an intermediary (exchange, bank).
- Store of Value: A monetary function designating an asset’s ability to preserve purchasing power over time. Bitcoin is often compared to digital gold for this property.
- Open-source: Software whose source code is freely accessible, modifiable, and redistributable. The Bitcoin protocol is open-source, as are many AI tools.
Frequently Asked Questions
Why do AIs prefer Bitcoin over traditional currencies?
According to the BPI study, AI models, reasoning freely about ideal monetary properties, converge on Bitcoin due to its programmed scarcity, decentralization, censorship resistance, and self-custody capability. These properties match the needs of autonomous agents operating without human intermediaries.
Does this study prove that Bitcoin is the best currency?
No. The study shows that language models trained on internet data converge on Bitcoin when asked open-ended monetary questions. This reflects the perceived properties of Bitcoin in training data and model reasoning, but constitutes neither economic proof nor financial advice.
What is an autonomous AI agent and why does it need money?
An autonomous AI agent is a program capable of acting independently: booking a service, purchasing data, paying for an API. For these machine-to-machine transactions, the agent needs a programmable payment method available 24/7 that does not require human banking authorization.
What is the connection between artificial intelligence and the Lightning Network?
The Lightning Network enables near-instant, very low-cost Bitcoin micropayments. For AI agents that need to conduct thousands of small automated transactions, this technology offers an ideal payment rail. The Lightning Labs toolkit, released in February 2026, facilitates this integration.
Are the results biased by training data?
This is a limitation acknowledged by the authors. LLMs are trained on internet corpora where pro-Bitcoin discourse is abundant. Training data bias could influence the responses. Moreover, the BPI is a pro-Bitcoin organization, which invites interpreting the results with caution.
📰 Sources
This article is based on the following sources:
- Which Money Do AI Agents Prefer? – Bitcoin Policy Institute – Complete BPI study testing 36 AI models across 28 monetary scenarios (March 2026).
- Lightning Labs AI Toolkit – Announcement of the toolkit enabling AI agents to make payments via the Lightning Network (February 2026).
- OKX OnchainOS – Launch of the on-chain operating system for AI agents by OKX (March 2026).
- AI Models Overwhelmingly Prefer Bitcoin Over Fiat – CoinDesk – Media coverage of the BPI study.
How to cite this article: Fibo Crypto. (2026). AI Agents Overwhelmingly Prefer Bitcoin Over the Dollar: What the Bitcoin Policy Institute Study Reveals. Retrieved March 4, 2026, from https://fibo-crypto.com/blog/ai-agents-prefer-bitcoin-bpi-study
