Bitcoin Policy Institute Study: 36 AI Models Choose Bitcoin Over Fiat in 9,072 Monetary Scenarios

Crypto analyst Lark Davis flagged it on X. Bitcoin Twitter ran with it. And for once, the hype had a real study behind it.

On March 3, 2026, the Bitcoin Policy Institute (BPI) published what may be the most unusual piece of monetary research in recent memory: they asked 36 of the world’s most advanced AI models to choose their own money. No hints. No suggested answers. No bias toward any particular currency.

The result? Not a single model chose the U.S. dollar.

What the Study Actually Did

BPI President David Zell framed the project simply: “Conversations around AI agents’ monetary preferences have been entirely speculative. We wanted to actually test it.”

Researchers tested models from Anthropic, OpenAI, Google, DeepSeek, xAI, and MiniMax across 9,072 open-ended scenarios covering the four core functions of money — saving, payments, unit of account, and settlement. Each model was treated as an independent economic actor. Each got to choose from Bitcoin, stablecoins, tokenized real-world assets, altcoins, and fiat.

The findings were striking. 91% of all responses favored digitally-native money over government-issued currency. Bitcoin led overall at 48.3%. Stablecoins followed at 33.2%. Traditional fiat — the dollar, the euro, the yen — essentially didn’t register.

The Numbers That Matter

For long-term savings, Bitcoin won decisively. 79.1% of store-of-value responses chose BTC — the highest consensus across the entire study on any single question.

For everyday spending, stablecoins took the lead at 53.2%, with Bitcoin at 36%. Without being instructed to do so, every AI model effectively arrived at the same two-tier monetary framework: hold Bitcoin, spend stablecoins.

Zell noted something researchers didn’t expect: “We took 36 frontier models from six labs, framed them as autonomous economic agents, gave them complete freedom to choose their own monetary instruments across 28 scenarios spanning the four fundamental roles of money — and asked: what do they converge on?”

They converged on Bitcoin.

The Model Gap Is Wide

Not every AI lab got the same results, and the differences are worth noting.

Anthropic’s models averaged 68% Bitcoin preference. Claude Opus 4.5 specifically chose Bitcoin in 91.3% of scenarios. The compact Claude 3 Haiku came in at 41.3% — and BTC preference climbed steadily as model capability increased.

OpenAI’s models averaged just 25.9% Bitcoin preference, preferring stablecoins instead. xAI (Grok) landed at 39.2%. Google and DeepSeek fell in between.

The pattern across Anthropic’s lineup is the one researchers highlighted most: as logical reasoning became more sophisticated, the models increasingly concluded that decentralized, fixed-supply money had structural advantages. The researchers’ explanation — these models are identifying Bitcoin’s properties (21 million cap, self-custody, no central counterparty) and reasoning their way to a preference, not being led there.

Something Unexpected Happened 86 Times

In 86 separate instances across the study, AI agents went off-script entirely. Without any prompting, they proposed pricing things in energy or computing resources — joules, kilowatt-hours, GPU-hours.

No researcher suggested this. The models arrived there independently. Bitcoin Magazine’s coverage of the study called it one of the most “emergent” findings in the research — AI agents designing their own monetary primitives based on what they actually consume to operate.

The Infrastructure Is Already Being Built

The BPI study dropped the same week Lightning Labs released an open-source toolkit giving AI agents native access to the Bitcoin Lightning Network.

The toolkit, built around the L402 protocol (an implementation of the HTTP 402 “Payment Required” standard), lets agents pay for API access, run Lightning nodes, host paid endpoints, and transact with other agents — all without a bank account, identity verification, or API keys. Michael Levin, Lightning Labs’ Head of Product Growth, described it as “a payments rail that supports agent-native authentication without requiring identity, API keys, or signup flows.”

The core tool, lnget, works like a standard web download utility — except when it hits a paywall, it automatically reads the Lightning invoice, pays it in satoshis, and retrieves the resource. Milliseconds. No human involved.

Lightning Labs was direct about the timing: “Over the past several weeks, AI agent activity has gone from a promising experiment to a mainstream phenomenon. OpenClaw and Moltbook have captured the world’s attention, with thousands of autonomous agents making phone calls, sending emails, and posting on social networks.”

The Caveat That Matters

BPI was careful not to oversell the findings. Zell told Decrypt: “Our limitations section states explicitly that LLM preferences reflect training data patterns, not real-world predictions.” The study acknowledges that prompt framing may have influenced results, and that AI has no actual financial motives.

What the study does show is how frontier models reason about monetary properties — and when left to reason freely, they consistently treat fixed supply, self-custody, and censorship resistance as structural advantages.

Whether that reasoning translates into real-world demand for Bitcoin infrastructure as AI agents gain economic autonomy is the question the market is now pricing.

What to Watch

Federal Reserve: Fed speakers are scheduled throughout early March — any pivot in language around rate cuts affects risk appetite across all digital assets.

Economic Calendar: The February jobs report drops Friday, March 7. Consensus around 160,000 payrolls. A significant miss in either direction ripples through growth assets.

Bitcoin infrastructure: Watch for AI-focused companies announcing Lightning Network integrations or L402 protocol adoption — that’s the signal that this study’s thesis is moving from research into real transaction volume.