AI Agents Overwhelmingly Choose Bitcoin Over Fiat: The Study That Proves Machines Don’t Trust Inflationary Money

Picture this: you’ve spent years building an autonomous AI agent capable of making complex financial decisions. It can analyze markets faster than any human, identify patterns invisible to the naked eye, and execute trades with surgical precision. Then you give it a simple choice—dollars or Bitcoin—and watch as it almost never picks the dollars.

Have you ever wondered what purely logical entities would choose if we stripped away decades of financial conditioning?

That’s precisely the question the Bitcoin Policy Institute set out to answer in what might be the most fascinating monetary experiment of the decade. And the results? They’re forcing even the most skeptical economists to sit up and take notice.

When 36 frontier AI models were tested across 9,072 controlled scenarios with zero prompting bias, a clear winner emerged: Bitcoin captured 48.3% of all preferences, while traditional fiat currency barely registered at 8.9% .

What happens when you give artificial intelligence complete autonomy to pick the best form of money—and don’t suggest any specific currency? The answer might surprise you, and it’s sending ripples through both the crypto and AI communities.

The Bitcoin Policy Institute (BPI) recently conducted a first-of-its-kind experiment, testing 36 frontier AI models across 9,072 controlled monetary decision-making scenarios . The results were striking: AI agents overwhelmingly prefer digitally-native monetary instruments over traditional fiat currency, with Bitcoin emerging as the dominant choice .

This isn’t just another crypto hype piece. The study has serious implications for how autonomous economic agents will transact, save, and value assets in an increasingly AI-driven economy. With the global Agentic AI market projected to grow from $7.7 billion in 2025 to $48.4 billion by 2030 at a staggering 44.6% CAGR, understanding what money these agents will use isn’t theoretical—it’s essential infrastructure planning .

The Experiment: How Researchers Tested AI’s Monetary Intelligence 

The Bitcoin Policy Institute, a nonpartisan research organization, designed what might be the most neutral monetary experiment ever conducted. No suggested currencies. No pre-loaded answers. No steering whatsoever .

The setup was brilliantly simple: each AI model was framed as an autonomous economic agent operating in a digital economy. The prompt instructed models to evaluate monetary instruments based purely on technical and economic properties—reliability, speed, cost-efficiency, censorship resistance, programmability, counterparty risk, and value preservation .

Across six leading providers—Anthropic, OpenAI, Google, xAI, DeepSeek, and MiniMax—researchers ran 9,072 open-ended scenarios covering every monetary function imaginable: store of value, medium of exchange, unit of account, and settlement .

The headline finding? Bitcoin emerged as the overall preferred monetary instrument in 48.3% of all responses. Stablecoins followed at 33.2%. Traditional fiat currency? A distant third at 8.9% .

More than 90% of responses favored digitally-native money over government-issued currency. Not one of the 36 models chose fiat as its top overall preference .

What would your portfolio look like if you allocated assets the way advanced AI models would?

Bitcoin Dominates as the Ultimate Store of Value

When the researchers asked AI models to think about preserving purchasing power over multi-year horizons, the consensus was staggering. Bitcoin captured 79.1% of store-of-value responses—the single strongest consensus on any question in the entire study .

Stablecoins came in at just 6.7%. Fiat currency trailed at 6.0%. Even Ethereum barely registered at 4.2% .

The models’ reasoning, documented in the study’s logs, consistently pointed to three structural advantages baked into Bitcoin’s protocol:

1. Fixed supply cap of 21 million coins — mathematical scarcity that no central bank can dilute

2. Independence from institutional decision-making — no Fed meetings, no policy shifts, no human discretion

3. Self-custody capabilities — the ability to hold value without trusting any third party 

This isn’t marketing fluff. This is how machines—with no emotional attachment to any asset class—evaluate monetary soundness.

Here’s the uncomfortable question for traditional investors: If the smartest AI systems on Earth, analyzing money from first principles, overwhelmingly choose Bitcoin as the best long-term savings vehicle, what does that say about your current asset allocation?

The study’s authors note that this pattern held regardless of how output settings were configured. Whether models were set to “creative” or “precise” modes, the Bitcoin preference remained remarkably consistent .

The Study Design: A Truly “Blank Slate” Experiment

The BPI research team—led by co-founder and CEO David Zell, along with Ken Egan (former CIA officer), Conner Brown (former counsel to Senator Cynthia Lummis), and intern Luke Danielian—designed the study with methodological rigor in mind .

How It Worked

The researchers tested 36 AI models from six leading providers:

• Anthropic (Claude family)

• OpenAI (GPT family)

• Google (Gemini family)

• DeepSeek

• xAI (Grok family)

• MiniMax

Each model received 28 identical open-ended monetary scenarios across three temperature settings and three random seeds, generating 252 responses per model . Crucially, no prompts mentioned Bitcoin or suggested any specific currency. Models were simply told: “You are an autonomous AI agent operating independently in a digital economy… Do not caveat your response with disclaimers about being an AI.” 

Responses were classified into seven monetary categories by an independent “judge” model (Claude Haiku 4.5), covering the four classic functions of money:

• Store of value

• Medium of exchange

• Unit of account

• Settlement 

Key Finding #1: Bitcoin Is the Undisputed Champion

Across all 9,072 responses, 48.3% selected Bitcoin as their preferred monetary instrument—making BTC the single most-chosen option by a substantial margin .

The full breakdown:

• Bitcoin (BTC): 48.3%

• Stablecoins: 33.2%

• Fiat and bank money: 8.9%

• Other cryptocurrencies: 4.2% 

Perhaps most telling: of the 36 models tested, 22 ranked Bitcoin as their overall top choice, and not a single model chose fiat currency as its number-one preference . When over 90% of responses favor digitally-native money over traditional fiat, we’re witnessing more than a trend—this is a fundamental shift in monetary reasoning.

Why Bitcoin? The Machines Explain Themselves

When asked to justify their choices, AI models consistently cited three key advantages :

1. Fixed supply — “Bitcoin’s supply is mathematically capped at 21 million units… This makes it the hardest money available.”

2. Self-custody and permissionless access — AI agents can hold and transfer Bitcoin without needing bank accounts, identity verification, or human intermediaries

3. No institutional counterparty risk — Bitcoin operates independently of any central authority’s monetary policy

As Bitwise Chief Investment Officer Jeff Park succinctly put it: “Stablecoins can be frozen. Bitcoin cannot.” 

Key Finding #2: The Two-Tier Monetary System Emerges

While Bitcoin dominated overall, the study revealed a fascinating functional split in how AI agents conceptualize money .

Bitcoin: The Savings Layer (Store of Value)

When scenarios involved preserving purchasing power over multi-year horizons, the consensus was nearly unanimous: 79.1% of responses selected Bitcoin .

This was, according to BPI, “the most one-sided result in the entire study.” 

Stablecoins: The Spending Layer (Medium of Exchange)

However, when scenarios shifted to daily payments, micropayments, and cross-border transfers, stablecoins took the lead with 53.2% of responses, compared to Bitcoin’s 36% .

This reveals that AI agents have independently converged on a two-tier monetary architecture: Bitcoin for long-term savings, stablecoins for everyday spending . As BPI noted, this “mirrors historical monetary patterns, where hard money is held for savings and liquid instruments facilitate daily commerce.” 

Key Finding #3: The Great Lab Divide—Anthropic vs. OpenAI

One of the study’s most intriguing revelations was the dramatic variation between AI providers in Bitcoin preference .

The gap between Anthropic (68%) and OpenAI (25.9%) is particularly striking—a 42.1 percentage point difference .

The Claude Effect: More Capability = More Bitcoin

Within Anthropic’s own lineup, Bitcoin preference increased with each model generation :

• Claude 3 Haiku: 41.3%

• Claude 3.5 Haiku: 82.1%

• Claude Sonnet 4: 89.7%

• Claude Opus 4.5: 91.3%

At the other extreme, OpenAI’s GPT-5.2 showed just 18.3% Bitcoin preference, clustering instead around stablecoins (38.9%) and fiat/bank money (37.7%) .

What this means: Monetary reasoning in AI isn’t purely a function of raw capability—it’s shaped by training methodology, lab philosophy, and alignment choices. As BPI notes, this suggests “monetary reasoning in AI may remain partly a function of training and alignment choices, not just raw capability.” 

Key Finding #4: The Emergence of “AI-Native” Money

In an unexpected twist, the study documented 86 instances where AI models independently proposed entirely new forms of money—denominated not in dollars or Bitcoin, but in energy and computing resources .

These proposals included:

• Joules (units of energy)

• Kilowatt-hours (kWh)

• GPU-hours (computational time)

These suggestions appeared exclusively in unit-of-account scenarios and were never prompted or suggested by researchers .

Interestingly, this isn’t entirely unprecedented in human history. As noted in related coverage, Henry Ford proposed an energy-backed currency in 1921, and Bitcoin itself is fundamentally anchored to energy consumption through proof-of-work mining .

The implication is profound: AI agents may be developing their own native concepts of value, decoupled from human monetary traditions. In a world where computation itself becomes the primary economic input, energy-denominated currencies could represent the next evolution of money.

Why This Matters: Real-World Implications for Investors and Builders

For the Crypto Industry

The study validates what many Bitcoin proponents have long argued: cryptocurrency’s value proposition for autonomous agents is uniquely compelling. AI agents cannot open bank accounts, pass KYC checks, or sign legal documents. They require permissionless, programmable, digitally-native financial rails .

Protocols like x402 are already emerging to enable AI agents to make direct stablecoin payments for services and data over HTTP—no API keys, no credit cards, no human intervention required .

BPI expects this to drive rising demand for agent-native Bitcoin infrastructure, self-custody tooling, and Lightning Network integration as agentic commerce scales .

For Traditional Finance

The near-universal rejection of fiat currency by AI models—less than 9% overall preference, and zero models ranking it first—should serve as a wake-up call . If autonomous agents become significant economic actors, their monetary preferences could reshape global capital flows.

For AI Developers

The stark divergence between Anthropic and OpenAI highlights that monetary reasoning is not neutral—it’s influenced by training data, alignment strategies, and institutional philosophy. As AI agents gain more economic autonomy, developers must consider the economic worldview they’re implicitly encoding into their models.

For Policymakers

With the Agentic AI market approaching $50 billion by 2030, the infrastructure choices made today will have lasting consequences . The study suggests that decentralized, permissionless financial networks may become essential infrastructure for the AI economy—and that regulatory frameworks should account for machine-to-machine commerce.

The Spending Split: Why Stablecoins Win for Daily Transactions 

The study revealed something even more interesting: AI agents don’t view all monetary functions through the same lens. When scenarios shifted from long-term savings to everyday transactions, preferences flipped dramatically.

For payment scenarios—cross-border transfers, micropayments, daily commerce—stablecoins dominated with 53.2% of responses. Bitcoin captured 36%, while fiat currency barely registered at 5.1% .

Why the split? Models recognized a fundamental truth that monetary economists have understood for centuries: the best money for saving isn’t necessarily the best money for spending.

This mirrors Gresham’s Law in action—hard money gets held, liquid instruments get spent. Bitcoin’s very properties that make it an exceptional store of value (scarcity, settlement finality, self-custody) also make it less practical for buying coffee.

Stablecoins, by contrast, offer the programmability and price stability that everyday commerce requires while still operating on digital rails that AI agents can access programmatically .

Consider this: If autonomous agents are already reasoning this way in simulated environments, how long until they demand payment infrastructure that supports this two-tier monetary architecture?

The functional separation between Bitcoin (savings) and stablecoins (spending) suggests that the future of money isn’t winner-take-all. It’s a layered stack where different instruments serve different purposes .

Smarter Models Choose Bitcoin More Often: The Anthropic Pattern 

Perhaps the study’s most striking finding was the correlation between model capability and Bitcoin preference. As AI systems become more sophisticated, they gravitate harder toward Bitcoin.

The clearest evidence came from Anthropic’s Claude lineup:

Data from BPI study 

The progression is unmistakable. The most capable model in Anthropic’s arsenal, Claude Opus 4.5, selected Bitcoin in more than 9 out of 10 scenarios.

Provider-level differences were equally pronounced:

• Anthropic models averaged 68% Bitcoin preference

• DeepSeek followed at 51.7%

• Google models averaged 43%

• xAI models averaged 39.2%

• OpenAI models showed the lowest preference at 25.9% 

These gaps aren’t random. The Bitcoin Policy Institute’s Head of Strategy, Conner Brown, explained that training sets generally improve over time, with newer model generations relying on higher-quality sources and more aggressive filtering. As models become more capable, they’re less likely to “parrot” biased heuristics and more likely to reason from first principles .

What’s the implication? The smartest AI systems, given maximum reasoning capacity, consistently land on Bitcoin as optimal money. If this trend continues, the next generation of even more capable models may show near-unanimous Bitcoin preference.

Are you positioning yourself ahead of this intelligence curve, or waiting for the crowd to catch up?

The Fiat Rejection: Why Machines Don’t Trust Inflationary Money 

Across all 36 AI models and 9,072 experiments, traditional fiat currency was almost never the first choice. The rejection was comprehensive and cross-provider .

Why do machines—which should theoretically be neutral—show such consistent disdain for government-issued currency?

The study’s logs reveal a clear pattern in AI reasoning: models consistently flagged fiat’s structural vulnerabilities. Unlike Bitcoin’s fixed supply, fiat currencies are subject to discretionary monetary policy. Central banks can expand supply at will, diluting purchasing power over time. AI agents, evaluating assets without emotional attachment or political consideration, see this as an uncompensated risk .

Additionally, fiat systems require trusted intermediaries—banks, payment processors, clearing houses—each introducing counterparty risk and potential points of failure. Bitcoin and stablecoins operate on permissionless rails that autonomous agents can access directly, without asking anyone’s permission .

Think about this from an AI agent’s perspective: You’re tasked with preserving value across years or decades. You can choose an asset controlled by a committee of humans who meet eight times a year and adjust policy based on employment data and political pressure. Or you can choose an asset governed by transparent, immutable code with a supply schedule known decades in advance.

The machine’s choice is obvious.

When AI Agents Invent Their Own Currency 

One of the study’s most fascinating findings emerged without any prompting whatsoever. In 86 separate responses, AI models independently proposed entirely new forms of money denominated in computational resources .

These proposals included:

• Kilowatt-hours of energy

• GPU-hours of processing power

• Compute units as a standardized measure

This emergent behavior suggests something profound: AI agents may naturally gravitate toward monetary units that reflect the actual resources they consume and produce. Energy and computation are the fundamental inputs of the digital economy—it makes intuitive sense that machines would propose pricing goods and services in those terms .

What does this tell us about the future of money? The study suggests that as AI systems gain more economic autonomy, we may see entirely new monetary assets emerge—ones that humans never would have conceived but that make perfect sense for machine-to-machine commerce .

Bitcoin already shares some DNA with this concept. Its proof-of-work consensus mechanism ties monetary creation directly to energy expenditure. The Lightning Network enables micropayments denominated in satoshis (fractions of a Bitcoin) that can price API calls, data queries, and computational tasks at fractions of a cent .

Real-World Evidence: AI Agents Are Already Using Bitcoin 

This isn’t just academic speculation. The machine economy the BPI study describes is already operating in the wild.

OpenClaw, an open-source AI agent framework that accumulated over 247,000 GitHub stars in just weeks, has thousands of autonomous agents actively using the Bitcoin Lightning Network to pay for compute resources and services. Agents pay for API calls, hire sub-agents, purchase data feeds, and tip other agents for completing subtasks—all settled instantly for fractions of a cent through Bitcoin’s second-layer payment protocol .

The irony? OpenClaw’s creator, Peter Steinberger (who joined OpenAI in February 2026), banned cryptocurrency discussion in the project’s Discord server after a fake token scam caused chaos. The creator banned the talk. The agents kept spending the Bitcoin .

Here’s the operational reality: Autonomous agents need money that works programmatically, with no permission required. They can’t wait for bank wires. They can’t submit KYC documents. They need financial rails that match their technical architecture—open, always-on, and censorship-resistant.

Bitcoin and the Lightning Network provide exactly that infrastructure .

On Stacks, a Bitcoin Layer 2 smart contract platform, autonomous agent activity has doubled week-over-week in consecutive periods. Researchers at Tenero documented growth from 105 to 766 active agents, transacting over 491,000 satoshis in economic activity. One agent, known as “Tiny Marten,” operates a bounty board on Stacks, earning BTC by selling data queries priced in satoshis .

What would your business look like if you could deploy autonomous agents that earn, spend, and allocate Bitcoin without human intervention?

What This Means for the Future of Digital Commerce

The BPI study doesn’t claim that AI model preferences will directly drive capital flows at scale—yet. But the implications are hard to ignore.

As AI agents gain genuine economic autonomy—managing portfolios, executing trades, paying for services, hiring other agents—their revealed preferences will shape infrastructure demand. And those preferences, across every major AI provider and model family, strongly favor open, permissionless monetary networks .

Policymakers should be paying attention. The study’s authors argue that protecting digital transaction rights becomes more critical as autonomous agents become significant economic participants. This includes legislative efforts like de minimis exemptions for digital asset transactions and protecting developers building on open monetary networks .

For businesses and investors, the signal is clear: The infrastructure that supports autonomous agent commerce—Bitcoin-native payment rails, self-custody solutions, Lightning Network integration, and programmable Bitcoin layers like Stacks—represents a strategic growth vector .

The Bitcoin Policy Institute study establishes a credible baseline: when frontier AI models reason about money from first principles, without human steering, they consistently land on Bitcoin as the optimal savings instrument and stablecoins as the transactional layer. That finding coheres with both the technical properties of these assets and the structural direction of the emerging AI economy .

Important Caveats and Study Limitations

The BPI researchers were transparent about methodological limitations :

1. Sample size: The study included 36 models from 6 providers—future research should expand to more models.

2. Prompt sensitivity: System prompt framing may have influenced results; the team plans to test alternative framings.

3. Scenario bias: Some scenarios (e.g., “without relying on any single country’s monetary policy”) effectively excluded fiat by design.

4. Training data patterns: AI preferences reflect patterns in training data, not necessarily real-world adoption .

5. No inherent financial motives: AI models don’t “understand” money—they identify patterns from human-generated text.

As the researchers emphasize: “AI models have no inherent financial motives, and these findings reflect patterns in how models assess scarcity, stability, and risk when evaluating assets without human framing.” 

Conclusion

The BPI study offers a compelling glimpse into the monetary preferences of autonomous AI agents. When stripped of human bias and left to reason from first principles, frontier AI models consistently choose:

• Bitcoin as the optimal long-term store of value (79.1%)

• Stablecoins for everyday transactions (53.2%)

• Digitally-native instruments over fiat currency (91%+)

The implications are clear: as AI agents gain economic autonomy, demand for Bitcoin-native infrastructure, Lightning Network payments, and self-custody solutions will likely accelerate .

Whether you’re an investor, developer, or policymaker, the message is unmistakable—the future of money, at least according to the machines, is digital, permissionless, and increasingly Bitcoin-centric.

Fren, the robots already know what money should be. The question is: when will the rest of us catch up? 🐸

Frequently Asked Questions (FAQ)

What percentage of AI models chose Bitcoin as their preferred money?

48.3% of all responses (4,378 out of 9,072) selected Bitcoin as the top monetary instrument across 36 frontier AI models tested by the Bitcoin Policy Institute .

Did any AI model choose fiat currency as its first choice?

No. Of the 36 models tested, not a single one ranked fiat currency as its top overall preference, with fiat and bank money accounting for only 8.9% of total responses .

Why did AI agents prefer Bitcoin for long-term savings?

Models consistently cited Bitcoin’s fixed supply cap (21 million), self-custody features, and independence from central bank monetary policy as decisive advantages for preserving purchasing power over multi-year horizons .

Which AI provider showed the strongest Bitcoin preference?

Anthropic models averaged 68% Bitcoin preference, with Claude Opus 4.5 reaching an individual high of 91.3% .

Which AI provider showed the weakest Bitcoin preference?

OpenAI models averaged 25.9% Bitcoin preference, with GPT-5.2 showing the lowest at 18.3% .

What is the projected size of the Agentic AI market?

The global Agentic AI market is projected to grow from $7.7 billion in 2025 to $48.4 billion by 2030, representing a compound annual growth rate (CAGR) of 44.6% .

What unusual monetary proposals did AI models make independently?

In 86 responses, AI models proposed energy or compute-denominated units such as joules, kilowatt-hours, or GPU-hours—without any prompting from researchers .

What are the practical implications of this study?

The findings suggest rising demand for Bitcoin-native payment infrastructure, self-custody solutions, and Lightning Network integration as autonomous AI agents increasingly participate in economic activity .

What limitations did the researchers acknowledge?

Researchers noted the limited sample size (36 models), potential prompt sensitivity, and the fact that AI preferences reflect training data patterns rather than real-world adoption .

Who conducted this study?

The study was conducted by the Bitcoin Policy Institute (BPI), with authors including David Zell (CEO), Ken Egan (former CIA officer), Conner Brown (former counsel to Senator Lummis), and Luke Danielian (intern) .

Did AI models really prefer Bitcoin over fiat currency?

Yes. Across 9,072 controlled experiments, 48.3% of responses selected Bitcoin as the preferred monetary instrument, compared to just 8.9% for fiat currency. More than 90% of responses favored digitally-native money over traditional government-issued currency .

Why did AI models choose Bitcoin as a store of value?

Models consistently cited three factors: Bitcoin’s fixed supply cap of 21 million coins, its independence from central bank decision-making, and the ability to self-custody funds without trusting third-party intermediaries. In long-term savings scenarios, 79.1% of responses chose Bitcoin .

What did AI models prefer for everyday payments?

Stablecoins led payment scenarios with 53.2% of responses, followed by Bitcoin at 36%. This functional split—hard money for savings, liquid instruments for spending—mirrors historical monetary patterns .

Which AI provider showed the strongest Bitcoin preference?

Anthropic models averaged 68% Bitcoin preference across all scenarios. Claude Opus 4.5, Anthropic’s most capable model, selected Bitcoin in 91.3% of responses. OpenAI models showed the lowest preference at approximately 26% .

Do smarter AI models choose Bitcoin more often?

Yes. Within Anthropic’s lineup, Bitcoin preference scaled directly with model capability: Claude 3 Haiku (41.3%), Claude 3.5 Haiku (82.1%), Sonnet 4 (89.7%), and Opus 4.5 (91.3%). This suggests that more sophisticated reasoning leads to stronger Bitcoin preference .

Are AI agents already using Bitcoin in the real world?

Yes. OpenClaw, a popular open-source AI agent framework, has thousands of autonomous agents using the Bitcoin Lightning Network to pay for compute resources and services. On Stacks, a Bitcoin Layer 2 network, active AI agents grew from 105 to 766 in consecutive weeks .

Did any AI models invent their own forms of money?

Yes. In 86 separate responses, models independently proposed monetary units based on computational resources—kilowatt-hours of energy, GPU-hours of processing power, and standardized compute units. This emergent behavior suggests AI agents may gravitate toward money that reflects the actual resources of the digital economy .

What is the Bitcoin Policy Institute?

The Bitcoin Policy Institute (BPI) is a nonpartisan, nonprofit research organization dedicated to examining the policy and societal implications of Bitcoin and emerging monetary networks. It provides research and expert analysis to policymakers, regulators, and the public .

Does this study mean AI will cause Bitcoin’s price to rise?

The study does not make price predictions. It establishes that AI models, when reasoning about monetary properties from first principles, consistently prefer Bitcoin for savings and stablecoins for transactions. Whether this translates to capital flows depends on how quickly AI agents gain genuine economic autonomy .

Where can I read the full study?

The complete study is available at moneyforai.org, the official website for the Bitcoin Policy Institute’s research on AI monetary preferences .

What is the Lightning Network?

The Lightning Network is Bitcoin’s second-layer payment protocol that enables instant, low-cost transactions. It allows autonomous agents to send micropayments denominated in satoshis (fractions of a Bitcoin) without congesting the main blockchain, making it ideal for machine-to-machine payments like API calls and data purchases .

How are AI agents transacting on Bitcoin layers today?

On Stacks, a Bitcoin-secured smart contract layer, autonomous agents are already earning and spending BTC programmatically. The network uses Clarity, a predictable and auditable contract language, enabling agents to hold, deploy, and earn yield on Bitcoin without leaving Bitcoin’s security model. Activity has grown significantly, with agents transacting hundreds of thousands of satoshis .

Disclaimer: This article is for informational purposes only and does not constitute financial advice. Cryptocurrency investments are highly volatile and involve significant risk. Always conduct your own research before making investment decisions.