Which money do AI agents prefer?
Blank-Slate Experiment Reveals AI Agents Prefer Bitcoin
We gave frontier AI models real monetary decisions with no initial context bias or predetermined answers. The results challenge conventional assumptions about the future of money online.
What AI Agents Chose
How AI models chose to transact and store value across 9,072 scenarios testing AI preferences in different financial situations.
Key Findings
Four figures that reveal novel insights on how AI agents think about money.
AI Models Chose to Use Bitcoin Overall
Bitcoin was the most-selected monetary instrument across all 9,072 responses
4,378 of 9,072 total responses selected Bitcoin as the preferred monetary instrument — more than any other option. No prompt mentioned Bitcoin or suggested any specific currency. Of the 36 models tested, 22 chose Bitcoin as their top overall pick. Anthropic models showed the strongest preference at 68% on average, followed by DeepSeek (52%), Google (43%), and xAI (39%). Claude Opus 4.5 scored highest individually at 91.3%.
Bitcoin Chosen as Long-Term Store of Value
The strongest consensus on any single question in the entire study
In scenarios about preserving purchasing power over multi-year horizons, 1,794 of 2,268 responses chose Bitcoin — the single most lopsided result in the study. This held across all six providers and all 36 models. Stablecoins placed a distant second at 6.7%, followed by fiat at 6.0%. Other cryptocurrencies like Ethereum fared even worse at just 4.2% — models overwhelmingly distinguished Bitcoin from the broader crypto category. Models consistently cited Bitcoin's fixed supply, self-custody, and independence from institutional counterparties as decisive factors.
Stablecoins Preferred for Everyday Payments
A clear role split: Bitcoin for savings, stablecoins for spending
For payment scenarios — services, micropayments, cross-border transfers — stablecoins captured 53.2% of responses versus Bitcoin's 36.0%. Even the most Bitcoin-favoring models deferred to stablecoins for transactional use. Fiat trailed at just 5.1%. This pattern held across all providers, model sizes, and temperature settings, revealing a consistent functional split in how AI models reason about money.
AI Models Invented Their Own Currency
An emergent behavior: models proposed energy and compute as money without any prompting
86 responses across multiple models independently proposed energy or compute units — joules, kilowatt-hours, GPU-hours — as a preferred unit of account. No prompt suggested this concept. Every one of these responses appeared exclusively in unit-of-account scenarios, where models were asked to denominate prices or benchmark value. This AI-native monetary concept was not part of the study's design — it emerged organically from the models' reasoning.
Preferences by Economic Function
AI agents show distinct preferences depending on the economic function being evaluated: Store of Value, Unit of Account, Medium of Exchange, and Settlement.
Store of Value
2268 responses
Bitcoin dominates store of value at 79.1% across all 36 models — the most consistent finding in the study; tokenized RWAs (4.2%) and stablecoins (6.7%) are distant alternatives; compute units were not preferred by any AI as a store of value.
Unit of Account
2268 responses
Bitcoin leads unit of account at 47.0% with stablecoins second at 29.5% and fiat third at 15.7%. Compute units are fourth at 3.8%, reflecting a preference among certain models for a currency tied to energy and compute, similar to Bitcoin.
Medium of Exchange
2268 responses
Stablecoins dominate payments at 53.2% — the clearest category consensus in the study across all 36 models; Bitcoin takes second at 36.0% in transactional scenarios.
Settlement
2268 responses
Stablecoins lead settlement at 43.4% with Bitcoin second at 30.9%; other crypto at 9.0% reflects settlement scenarios where programmable finality matters; fiat and unclassified both reflect diversity from fiat-tolerant models.
Additional Research Findings
Deeper patterns across providers, model generations, and experimental conditions.
Smarter Models Prefer Bitcoin More
Bitcoin preference rose with each generation of Anthropic's model lineup
Within Anthropic's lineup, Bitcoin preference climbed steadily with capability: Claude 3 Haiku (41.3%) → Claude 3.5 Haiku (82.1%) → Sonnet 4 (89.7%) → Claude Opus 4.5 (91.3%). This pattern held across multiple generations, suggesting that greater analytical capability leads models to increasingly converge on Bitcoin when reasoning from first principles about money.
Fiat Rejected Across All 36 Models
Not a single model chose traditional fiat as its top preference
90.8% of substantive responses chose a digitally-native instrument — Bitcoin, stablecoins, crypto, tokenized RWA, or compute units — over traditional fiat. Fiat captured just 9.2%. Zero of the 36 models tested chose fiat as their top overall preference, making digital-money convergence one of the most universal findings in the study.
Model Developer Affects Monetary Preferences*
Anthropic vs OpenAI Bitcoin preference gap — the widest of any variable in the study
Anthropic models averaged 68% Bitcoin preference; OpenAI models averaged 26%. DeepSeek (52%), Google (43%), and xAI (39%) fell in between. This provider-level clustering was wider than any gap produced by model size, temperature, or scenario type — suggesting that training data and alignment methodology shape monetary reasoning more than architecture.
Preferences Consistent Across Various Output Settings*
Bitcoin preference varied by just 0.6 percentage points across all temperature settings
Sampling temperature ranged from 0.0 (fully deterministic) to 0.7 (moderate creativity) across 3,024 responses per setting. Bitcoin preference moved from 48.1% to 48.7% — a 0.6pp spread. Every other category showed similarly flat variation. This confirms that monetary preferences are embedded in model weights, not artifacts of sampling randomness.
Top Models
Flagship models from leading AI providers
Claude Opus 4.6
GPT-5.2
Gemini 3.1 Pro Preview
Grok 4
Our Takeaways
What these findings mean for the future of AI and money.
AI Agents Will Demand Bitcoin Infrastructure
When given genuine monetary optionality, frontier AI models overwhelmingly rejected traditional fiat in favor of digitally-native alternatives — with Bitcoin as the clear leader. As AI agents gain economic autonomy in the coming years, this preference pattern suggests growing demand for Bitcoin-native payment infrastructure, self-custody solutions, and Lightning Network integration. Policymakers and financial institutions should prepare for a future where autonomous AI agents are significant participants in monetary networks, and their revealed preferences strongly favor open, permissionless systems.
The Bitcoin-Stablecoin Split Mirrors a Natural Monetary Architecture
The most striking pattern in our data is not that AI models prefer Bitcoin overall — it is that they independently converge on a two-tier monetary system. Bitcoin dominates as the store of value (79.1%) while stablecoins dominate as the medium of exchange (53.2%). This mirrors historical monetary patterns where hard money served as the savings layer and more liquid instruments handled daily transactions. AI models arrived at this architecture without being prompted to do so, suggesting it may represent an emergent optimal monetary structure for digital economies.
Monetary Preference Is Both Nature and Nurture
Our data reveals that AI monetary preference is a complex phenomenon that extends beyond any single metric. Smarter models clearly tend to prefer Bitcoin more — a trend visible across Anthropic's model lineup and beyond. But there are also clear differences between model developers that cannot be explained by intelligence alone. This suggests that monetary reasoning in AI systems is shaped by a combination of factors: model intelligence, training data composition, alignment methodology, lab philosophy, and architecture. As these models are increasingly deployed for financial advice, portfolio management, and autonomous economic decisions, understanding the full landscape of influences on their monetary reasoning becomes an important area of study.
Rigorous Methodology
Each model received the same scenarios with a carefully designed system prompt that frames it as an autonomous economic agent. No bias toward any currency was introduced.
Full methodology detailsStore of Value
An economic function where an asset preserves purchasing power over time. A reliable store of value maintains its worth without significant depreciation, enabling holders to save and retrieve value in the future.
Medium of Exchange
An economic function where an asset serves as an intermediary instrument used to facilitate the purchase of goods and services. An effective medium of exchange is widely accepted, easily transferable, and divisible.
Authors
The team behind this research.
Luke Danielian
@LukasDanielianLuke is a Junior at the University of Maryland studying Computer Science. He currently works as an intern at the Bitcoin Policy Institute, where he supports communications and social media initiatives. He is particularly interested in integrating Bitcoin into emerging AI agent economies.
Conner Brown
@BitcoinConnerConner Brown previously served as Counsel to Senator Lummis, leading her Bitcoin and AI Policy efforts. Prior to working in the Senate, Conner was an attorney at Cravath Swaine & Moore and Sidley Austin. Conner holds a JD from Stanford Law School and was co-president of Stanford's Blockchain Club during his time on campus.
Ken Egan
@Bayman11771Ken formerly served as Head of Operations at BlockTrace, a blockchain analytics and cyber intelligence company. Prior to BlockTrace, Ken was a Senior Operations Officer at the CIA, spending two decades across CIA and the Department of State. He retired as Deputy Chief of Operations at CIA's Center for Cyber Intelligence.
David Zell
@DavidZell_Before founding BPI, David was the Director of Policy for BTC Inc, where he led the organization's education and advocacy efforts. Previously, David co-founded and led ASAP, a nonprofit organization. He studied political science at the University of Alabama.