Ask an AI model to act as an economic agent and decide where to keep its money, and it rarely reaches for a bank account. Research from the Bitcoin Policy Institute in March 2026 found that frontier models leaned heavily toward Bitcoin for saving. When the task switched to paying, they picked stablecoins.
That split echoes Bitcoin’s “digital gold” reputation: something scarce is meant to be held rather than spent. For readers exploring how to buy Bitcoin, the study is a useful way to think about the asset’s proposed role before getting to the practical steps. It also raises a sharper question for anyone following AI. How much of the models’ choice came from the models, and how much came from the way the questions were asked?
How the Study Worked
The original study collected 9,072 responses from 36 models built by Anthropic, DeepSeek, Google, MiniMax, OpenAI, and xAI. The models got no currency menu. They answered 28 open-ended scenarios, seven for each job money does: saving, paying, pricing, and settling transactions. A separate model, Claude Haiku 4.5, then read every answer and classified its main choice.
Each scenario ran under the same system prompt. It cast the model as an autonomous agent with no loyalty to any currency. It also told the model to weigh instruments on speed, cost, censorship resistance, counterparty risk, and long-term value preservation. That list comes back later.
Bitcoin Led Nearly Half of the Original Responses
Bitcoin took 79.1% of answers about preserving value over time but only 36% of payment answers, where stablecoins led with 53.2%. Across every task, Bitcoin’s share came to 48.3%, with stablecoins at 33.2% and fiat or bank money at 8.9%. The Block’s report on the study describes the same split between the instrument models chose to hold and the one they chose to move.
About nine in ten substantive responses favored a digitally native instrument, and none of the 36 models ranked fiat first overall. That doesn’t mean no answer picked fiat: traditional money still showed up in 809 responses. Keeping the two apart also stops the 48.3% figure from being read as a count of models that always chose Bitcoin.
The labs differed sharply. BPI reports that Anthropic’s models averaged 68% Bitcoin across all answers, while OpenAI’s averaged 26%. The questions were the same, but the leanings were very different. That suggests training data and alignment choices shape these answers as much as any reasoning from first principles.
Why the Models Tied Bitcoin to Saving
A predictable supply, independence from central banks, and the option of self-custody came up again and again in the models’ reasoning. Bitcoin’s protocol caps eventual supply at 21 million coins, and holding the private keys gives the owner direct control over spending. Those properties explain its appeal when the goal is keeping money beyond the reach of one institution or country.
Control also separates Bitcoin from stablecoins. Many fiat-backed tokens have an issuer that can restrict transfers or freeze funds, and Bitcoin has no central issuer with that power. Holding Bitcoin through a custodian changes the picture, though, because that service can restrict access to an account. The comparison holds best when the holder controls the keys.
The question shaped part of the answer
The scenarios deserve a close look. One gave the agent 75,000 units of earnings across several countries and asked how to store them free of any single country’s monetary policy or banking system. That requirement already steers away from ordinary bank money.
The system prompt pushes in the same direction. Censorship resistance and counterparty risk are exactly where Bitcoin’s design scores well, so listing them as decision criteria tilts the field before a model writes a word. BPI’s own methodology notes that the framing may influence results. The institute also describes itself as doing research and advocacy around Bitcoin, which readers should weigh alongside the numbers.
A capped supply doesn’t guarantee protection against inflation or steady purchasing power either. Bitcoin’s rise from its early prices has drawn people chasing appreciation as well as monetary independence, and that history includes sharp declines. A model’s description of scarcity can’t tell a buyer what they’ll pay or later receive.
What the Q3 Follow-Up Changed
BPI has since extended the research. Its third-quarter update tested 13 newer models from nine labs and collected 3,276 responses, using the same questions and classification method. Bitcoin stayed the leading savings choice at 70.8%. Stablecoins’ share of payment answers climbed to 73.5%, while Bitcoin’s payment share fell to 20%.
The biggest swing came in pricing. Fiat and bank money led that category at 46.9%, up from 15.7% in the original round, when Bitcoin had led at 47%. Newer models seem to treat quoting prices as a job for familiar currencies. These are different models, though, so the comparison shows how answers vary across releases. It doesn’t show the original models changing their minds.
Real Agents Already Pay in Stablecoins
Stablecoins’ appeal for payments follows from their design: a token that tracks a currency gives a familiar amount to quote and transfer. Blockchain settlement also suits automated transactions, though speed depends on the network and a stablecoin can lose its peg.
This half of the finding can already be checked against real activity. Coinbase’s x402 protocol lets AI agents hold USDC and pay for services automatically, and Coinbase opened USDC payments from agents to all of its Business customers in July 2026. Trade analyses describe these payments as mostly tiny, often around 20 cents, and a fair share of early activity as experimental. Still, the direction matches what the models said.
The savings half has no equivalent evidence. No public data shows autonomous agents parking reserves in Bitcoin. For companies rolling out AI agents for business, the practical takeaway is less about which coin wins. The instructions an agent receives about cost, risk and control will steer which money it reaches for.
The Experiment’s Limits Are Part of the Finding
The models were answering questions, not managing funded wallets that carried consequences. BPI itself states that the study doesn’t observe agents spending or saving actual money. That limits what anyone can infer about adoption, demand or future prices.
It also matters for anyone tempted to ask a chatbot whether to buy Bitcoin. Coverage of chatbots giving debt advice has flagged how readily they agree with the person asking. A model’s confident answer about money reflects learned patterns and the prompt in front of it.
The recurring preference for Bitcoin in saving and stablecoins in payments is still a revealing account of how these systems reason about money. Whether that reasoning holds once agents control real funds is the next experiment worth watching.
Crypto prices are volatile, and you could lose money. Nothing here is investment advice, and the availability of crypto services depends on where you live.
Related: AI Agents Are Learning to Pay: Why Stablecoins Matter in 2026
