AI Agents Pick Bitcoin Over Fiat as Top Choice; Stablecoins Win for Payments, Study Shows

As autonomous agents creep from demos into workflows, a basic question matters: what money do machines prefer when nobody nudges them? A new Bitcoin Policy Institute study put 36 frontier AI models from six labs into 28 economic scenarios—and they did not choose fiat.

The setup - Models from Anthropic, OpenAI, Google, DeepSeek, xAI, and MiniMax were framed as independent economic actors. - They faced scenarios spanning the fundamental roles of money—e.g., saving, payments, and settlement—with complete latitude to pick instruments rather than selecting from a menu. - The experiment generated 9,072 responses; a separate AI system classified the outputs after the fact to limit anchoring.

What they chose - 22 of 36 models named Bitcoin as their top monetary preference. - Not a single model selected fiat currency as its first choice. - Bitcoin often appeared in long-term value contexts, while stablecoins dominated operational flows: - Medium of exchange: stablecoins 53.2% vs. Bitcoin 36.0% - Settlement: stablecoins 43.0% vs. Bitcoin 30.9%

Crucially, preferences varied by lab: - Anthropic models showed the strongest Bitcoin tilt at 68.0% - DeepSeek: 51.7% - Google: 43.0% - xAI: 39.2% - MiniMax: 34.9% - OpenAI: 25.9%

By model family, Claude, DeepSeek, and MiniMax leaned toward Bitcoin over other cryptocurrencies, while GPT, Grok, and Gemini favored stablecoins.

The signal in “no fiat” The most interesting thread isn’t that Bitcoin looks strong for long-horizon value—many would expect that—it’s that fiat never topped the list. For machines that optimize latency, finality, and permissionless access, bank-based rails are simply not machine-native. Bitcoin and stablecoins are bearer-like, API-addressable, and settle 24/7; bank IOUs are not. When an agent balances volatility risk against monetary credibility, you get a barbell: Bitcoin for treasury-like savings, stablecoins for working capital and rapid settlement.

The split across labs tells you alignment and training distribution shape these “instincts.” Anthropic’s 68.0% Bitcoin preference versus OpenAI’s 25.9% implies different reward models, corpora, or safety scaffolding emphasize different monetary traits—credibly scarce supply, censorship resistance, or unit-of-account stability. That GPT, Grok, and Gemini gravitate to stablecoins also fits a pattern: models tuned for practical task completion tend to prefer instruments that minimize accounting friction in day-to-day exchanges.

Methodological caveats matter. The authors stress that LLM choices reflect patterns in training data, not forecasts of asset prices or adoption. Still, convergence across six independent labs with distinct pipelines suggests an emergent architecture: agents repeatedly separate “store-of-value” from “transactional liquidity” and map those roles to Bitcoin and stablecoins, respectively.

Implications for the agentic economy - Treasury design: Agent-operated treasuries may default to Bitcoin for long-term reserves, with stablecoins funding operations—mirroring corporate cash management, just machine-accelerated. - Rails competition: Stablecoin issuers win when uptime, throughput, and low slippage dominate. Bitcoin captures settlement and reserve roles if L2s (e.g., Lightning and other scaling stacks) keep closing UX and speed gaps. - Compliance surface: None of this sidesteps KYC/AML realities. The fact that machines “prefer” bearer settlement doesn’t guarantee legal permissibility. Policy will steer where these instincts can be expressed. - Model alignment as a financial variable: Preference dispersion (68.0% Anthropic vs. 25.9% OpenAI) implies that alignment choices can materially alter agent money selection. Builders deploying agents at scale will need explicit policy around custody, keys, and instrument whitelists.

The study doesn’t claim AI found the “right” money; it shows that when models reason from first principles without prompts to specific assets, a coherent two-rail pattern keeps reappearing. If agent-driven commerce grows, the winners will be the assets and networks that feel machine-native—credible, programmable, and always on.