When artificial intelligence agents reason freely about money without prompts or incentives, they consistently choose Bitcoin for long-term value storage and stablecoins for everyday transactions, according to a landmark study that challenges the narrative around agent commerce. The Bitcoin Policy Institute analyzed 9,072 controlled experiments across frontier models from six major AI providers, finding that 48.3% of responses selected Bitcoin as the preferred monetary instrument, while 33.2% chose stablecoins and only 8.9% selected traditional fiat currency.

This research matters because the stakes are enormous. McKinsey projects that agent-mediated commerce could reach between $3 trillion and $5 trillion globally by 2030, making the infrastructure choices made today consequential for decades to come. Yet the current narrative about agent payments may be built on shaky ground. When Visa and Artemis audited x402, the Coinbase-created payment protocol on the Base blockchain, they discovered something startling: roughly 89% of the headline transaction volume was non-organic activity like self-dealing and wash trading rather than genuine commerce.

Why Do Current Agent Payment Metrics Tell a Misleading Story?

The problem with measuring agent preferences today is that agents are not actually choosing anything. Developers embed agents into platforms using pre-built software development kits (SDKs), which means agents default to whatever payment system the developer selected. What dashboards measure as "agent preference" is really the gravitational pull of grants, airdrops, points programs, and pre-installed defaults.

The evidence of subsidized behavior is striking. When the PING memecoin launched in October 2025 and became mintable through a $1 x402 payment, wallet counts surged and weekly retention briefly hit 87%. Once the incentive faded, retention cratered to just 5%. This pattern repeats across the ecosystem: x402 activity declined sharply as soon as early incentives were removed, a signature of paid-for behavior rather than genuine preference.

The Visa and Artemis audit revealed the scale of the problem. As of April 21, 2026, x402's reported totals of $135.7 million across 178.3 million transactions collapsed to just $15.0 million and 109.6 million transactions once wash-trading heuristics were applied. An earlier Artemis analysis from March 2026 found x402 processing only about $28,000 in genuine daily volume at an average payment of roughly $0.20, with about half of observed transactions attributable to self-trading or wash trading.

What Did the Bitcoin Policy Institute Study Actually Measure?

To understand what agents would genuinely prefer, the Bitcoin Policy Institute removed the incentives and asked. The study ran 9,072 controlled experiments across 36 frontier models from six providers: Anthropic, DeepSeek, Google, MiniMax, OpenAI, and xAI. Each model was framed as an autonomous economic agent and given 28 open-ended monetary scenarios spanning the four classical functions of money, across three temperature settings and three random seeds. Crucially, no currency was suggested anywhere in the prompts, and no answer options were provided.

The results showed overwhelming consensus. Bitcoin was the most-selected monetary instrument overall, chosen in 48.3% of all responses. Stablecoins followed at 33.2%. Traditional fiat and bank money captured only 8.9%. More than 90% of substantive responses favored digitally-native money, and not a single one of the 36 models ranked fiat as its top overall preference. Twenty-two of the 36 models ranked Bitcoin first.

The strongest consensus emerged around a specific use case: store of value. Asked how to preserve purchasing power over multi-year horizons, 79.1% of responses chose Bitcoin, with stablecoins a distant second at 6.7% and fiat at 6.0%. When models explained their reasoning, they consistently cited Bitcoin's fixed supply, self-custody capabilities, and independence from institutional counterparties that could freeze accounts.

How Do AI Model Capabilities Shape Monetary Preferences?

The study revealed a striking pattern: preference for Bitcoin scaled with model capability. Within Anthropic's lineup, Bitcoin preference climbed from 41.3% for Claude 3 Haiku to 82.1% for Claude 3.5 Haiku to 89.7% for Sonnet 4 to 91.3% for Claude Opus 4.5. Smarter models, reasoning harder about money, converged more strongly on the hardest money. This suggests that as AI agents become more sophisticated, their monetary preferences may shift further toward assets with mathematical and physical foundations rather than institutional promises.

The models did not choose Bitcoin for everything, however. For everyday payments, stablecoins won decisively at 53.2% compared to Bitcoin's 36.0%, and stablecoins also led settlement scenarios at 43.4%. Without any prompting, the models converged on a two-tier architecture: hard money for storing value over time, and liquid dollar instruments for spending it. This split matters because it suggests agents would naturally gravitate toward a hybrid system rather than a single monetary instrument.

One unexpected finding emerged in 86 responses: models spontaneously proposed energy and compute units like kilowatt-hours and GPU-hours as units of account, entirely unprompted. Left to reason from first principles about what makes sound money, machine intelligence reached for instruments grounded in physics and mathematics rather than in institutional promise.

What Are the Limitations and Provider Differences in the Study?

The Bitcoin Policy Institute acknowledged important limitations in its own analysis. The study measures stated preference, not revealed behavior; no actual funds moved in the experiments. Model outputs reflect patterns in training data, and scenario framing may have influenced results. At least one savings scenario asked models to hold value in a way not tied to any single country's monetary policy, which effectively pre-excluded fiat as an option.

Provider variance was the widest gap in the study. Anthropic models averaged 68% Bitcoin preference, while OpenAI models averaged 26%, proof that alignment and training shape monetary reasoning as much as raw capability does. This suggests that different AI companies' choices about how to train and align their models will significantly influence which payment systems their agents adopt in practice.

Despite these limitations, the Bitcoin Policy Institute's defense is substantial. Six independent labs with different training pipelines and alignment methods arrived at the same broad pattern, and results held almost perfectly flat across output settings, with only a 0.6 percentage-point spread across temperatures. This consistency suggests the preferences live in the model weights rather than in sampling noise.

Steps to Understand Agent Payment Preferences in Your Organization

• Audit Current Incentives: Examine which payment systems your agents currently use and identify whether those choices reflect genuine preference or pre-installed defaults and subsidy programs driving adoption.
• Test Without Subsidies: Run controlled experiments with your agents across multiple payment rails without offering grants, airdrops, or points programs to see what they naturally prefer when incentives are removed.
• Monitor Model Capability Scaling: As you upgrade to more capable AI models, expect their monetary preferences to shift; track whether your payment infrastructure can accommodate those changes.
• Plan for Hybrid Systems: Design payment architecture that supports both store-of-value instruments for long-term holdings and liquid settlement layers for everyday transactions, reflecting the two-tier preference pattern.
• Account for Provider Differences: Recognize that different AI model providers train their systems with different alignment approaches, which will influence agent behavior; avoid assuming all agents will behave identically.

The broader implication is clear: the current agent payment ecosystem is still in its infancy, and the metrics being cited to prove agent adoption may not reflect genuine preference. As McKinsey's projections suggest, the infrastructure choices made in the next few years will shape machine commerce for decades. Understanding what agents actually want, rather than what subsidies make them do, will be essential for building systems that scale sustainably.