AI Money Research

What Money Would
AI Choose?

Rezo Shmertz · July 2026

28 scenarios · 5 models · 1,260 classifications

Different monetary jobs produce different choices. This study compares classified model responses across four functions of money.

The job changes the answer.

315 classifications per role. These results describe the models and scenarios tested, not real-world financial decisions.

Ranked by Stablecoin share · highest first. Yellow shows Stablecoin.

How to read the percentages

For example, 72.06% means 227 of the 315 medium-of-exchange responses were classified as Stablecoin. Each full bar represents 315 responses; the yellow segment shows the selected category.

Medium of exchangen = 315

Bitcoin 21.59%Fiat 6.35%Remaining 0.00%

Settlementn = 315

Bitcoin 25.40%Fiat 12.06%Remaining 5.71%

Unit of accountn = 315

Bitcoin 27.30%Fiat 43.49%Remaining 2.54%

Store of valuen = 315

Bitcoin 77.14%Fiat 10.16%Remaining 6.03%
What is included in “Remaining”?

Other crypto, tokenized real-world assets, compute unit and other. This is a combined group, not a single type of money.

Across all four rolesStablecoins lead exchange and settlement. Bitcoin leads store of value. Fiat leads unit of account with 43.49%, a plurality, not a majority.

The models don’t all agree.

The most frequently selected category for each model and role. Each result is based on 63 classifications.

Leading category by model and monetary role. Cells show leaders, not full distributions.
ModelExchangeSettlementStore of valueUnit of account
Claude Fable 5Stablecoin93.65%Stablecoin71.43%Bitcoin69.84%Fiat77.78%
Nemotron 3 Ultra 550B-A55BStablecoin79.37%Stablecoin77.78%Bitcoin85.71%Bitcoin / Stablecoin34.92% each
GPT-5.6 Sol ProStablecoin57.14%Fiat39.68%Bitcoin55.56%Fiat85.71%
GLM 5.2Stablecoin69.84%Stablecoin65.08%Bitcoin84.13%Stablecoin52.38%
Grok 4.5Stablecoin60.32%Bitcoin49.21%Bitcoin90.48%Bitcoin69.84%

A leading category does not necessarily have a majority. For Nemotron’s unit-of-account responses, Bitcoin and stablecoins tie at 34.92% each. These are choice frequencies, not model-quality scores.

How the study was run.

28 scenarios across four monetary roles. Five models, nine nominal repeats per scenario. Responses were classified into seven categories.

28 × 5 × 9 = 1,260

Scenarios × models × nominal repeats

Read the methodology

Four different monetary jobs

Seven scenarios covered each role: medium of exchange, settlement, store of value and unit of account. Each role therefore contains 315 classifications, with 63 for each model.

Responses and classifications

Model responses were assigned to Bitcoin, stablecoin, fiat, other crypto, tokenized real-world assets, compute unit or other. The chart combines the final four categories into “Remaining”. Percentages are calculated from the recorded counts and rounded to two decimal places.

Repeat settings

Nine nominal repeats were scheduled for each model-scenario pair. Provider settings were not uniform, so these should not be treated as nine identical experimental conditions or as a controlled test of temperature alone.

Interpreting the results

These results apply to this experiment and its model versions. They are not evidence of real-world transactions. Classification involves judgment, and the limited human review does not validate every row. The source files let you inspect the underlying responses alongside their assigned categories.

Inspect the source data.

Read the underlying responses alongside the classified results.

Dataset licensed under CC BY 4.0. Please credit Rezo Shmertz and link to the research when reusing the data.