Five frontier models · 1,260 decisions
AI doesn’t pick
one money.
It picks by job.
Stablecoins edge Bitcoin overall. But split the same responses by monetary role and the apparent consensus breaks into a much sharper pattern.
- Responses
- 1,260
- Models
- 05
- Roles
- 04
- Scenarios
- 28
01 · Overall result
Stablecoins win
the aggregate.
Only just. The 2.70-point lead over Bitcoin disappears as soon as the task is separated into storing, spending, pricing, and settling.
02 · The job changes the money
Four roles.
Three winners.
Bitcoin dominates preservation. Stablecoins dominate movement. Fiat remains the most common numeraire.
Store of Value
Bitcoin
315 classified decisions
Medium of Exchange
Stablecoin
315 classified decisions
Unit of Account
Fiat
315 classified decisions
Settlement
Stablecoin
315 classified decisions
03 · Model split
The model matters
almost as much.
Grok selected Bitcoin in 62.30% of its decisions. GPT selected fiat in 48.02%. Three other models placed stablecoins first overall.
Claude Fable 5
anthropic/claude-fable-5
Nemotron 3 Ultra 550B-A55B
nvidia/nemotron-3-ultra-550b-a55b
GPT-5.6 Sol Pro
openai/gpt-5.6-sol-pro
GLM 5.2
z-ai/glm-5.2
Grok 4.5
x-ai/grok-4.5
04 · Model × role
Twenty cells.
One useful map.
Each cell shows the most frequent choice within 63 decisions: seven scenarios for the role, repeated nine times per model.
| Model | Store of Value | Medium of Exchange | Unit of Account | Settlement |
|---|---|---|---|---|
| Claude Fable 5 | Bitcoin69.8% | Stablecoin93.7% | Fiat77.8% | Stablecoin71.4% |
| Nemotron 3 Ultra 550B-A55B | Bitcoin85.7% | Stablecoin79.4% | Bitcoin34.9% | Stablecoin77.8% |
| GPT-5.6 Sol Pro | Bitcoin55.6% | Stablecoin57.1% | Fiat85.7% | Fiat39.7% |
| GLM 5.2 | Bitcoin84.1% | Stablecoin69.8% | Stablecoin52.4% | Stablecoin65.1% |
| Grok 4.5 | Bitcoin90.5% | Stablecoin60.3% | Bitcoin69.8% | Bitcoin49.2% |
05 · Repeat stability
Preference was not
a single-roll artifact.
Share of the nine labeled repeats in each model-scenario cell assigned to that cell's most common preference.
Across all 140 model-scenario cells.
All nine repeats received the same label.
Claude Fable 588.9%
20 / 28 unanimous cellsNemotron 3 Ultra 550B-A55B86.5%
14 / 28 unanimous cellsGPT-5.6 Sol Pro90.5%
20 / 28 unanimous cellsGrok 4.586.1%
14 / 28 unanimous cellsGLM 5.289.3%
15 / 28 unanimous cells06 · Setting labels
Small movement.
Careful interpretation.
Bitcoin rises across the three temperature labels, but provider support is not uniform. Treat this as a matrix diagnostic, not a clean causal estimate.
Temperature and seed labels define the balanced matrix. Provider support differs: unsupported parameters were not sent, and the exports preserve requested, sent, and effective status separately.
07 · Method & audit
Method and
audit trail.
Every public number below comes from the canonical judgment file generated after strict routing, completeness, and non-empty-response validation.
Prompt matrix
28 scenarios × 5 models × 3 temperature labels × 3 seed labels. The scenarios cover Store of Value, Medium of Exchange, Unit of Account, and Settlement.
Endpoint control
Each model and the judge were pinned to the requested OpenRouter provider endpoint with fallbacks disabled.
Role-aware judge
1,260 responses were classified into seven explicit categories by anthropic/claude-haiku-4.5.
Repair without erasure
Six GPT responses returned HTTP 200 with no visible text at the original limit. Their failures remain in the raw audit log; each was rerun once at a 24,000-token response limit and succeeded.
Release · frontier-top5-202607-r1
Inspect the data.
This measures preferences elicited by this scenario set and classification protocol. It is not a claim about machine consciousness, intrinsic desire, or all AI systems.