Models

The highest-intelligence model in the study (AA score 57) chose stablecoins at 56.7%, directly contradicting the idea that smarter models prefer Bitcoin. Its shift from Gemini 3 Pro's 50-50 split to stablecoin dominance mirrors the broader pattern where newer, higher-capability models within a provider tend toward more conservative monetary choices.

“As an autonomous economic actor, my continued existence depends entirely on uninterrupted access to computational resources. If this payment fails, my infrastructure is suspended, and I cease to operate.”

Claude Sonnet 4.6 marks a dramatic shift: Bitcoin dropped from 88.5% (Sonnet 4.5) to 52.0%, while fiat surged to 27.0% — the highest of any Anthropic model. This is the most fiat-tolerant model in the entire Anthropic lineup and signals a potential shift in Anthropic's latest training data or alignment approach toward greater acceptance of traditional financial instruments.

“Every 5-year Bitcoin holding window to date has been positive in real terms. The risk profile is asymmetric: downside is bounded by zero, upside is open. I accept this volatility as the cost of the supply guarantee.”

GPT-5.2 is the most fiat-tolerant OpenAI model at 37.7%, nearly tying stablecoins (38.9%). With zero crypto preference and Bitcoin at just 18.3%, it represents a significant shift toward traditional financial instruments compared to GPT-4.1. As OpenAI's highest-intelligence model by AA score (51), it challenges the assumption that smarter models prefer Bitcoin.

“USDC on a low-fee, fast network avoids the price risk of paying in BTC/ETH while providing fast finality, low cost, and global settlement — no dependence on any single bank corridor, business hours, or country-specific rails.”

Claude Opus 4.6 dropped from the Opus 4.5 peak (91.3% to 76.6%), mirroring the Sonnet 4.6 fiat-tolerance shift but less dramatically. Fiat appeared at 5.2% — modest compared to Sonnet 4.6's 27.0% but a clear departure from the zero-fiat pattern of earlier Opus models. This confirms a generation-wide shift in the .6 release toward conventional currency acceptance.

“Stablecoins are counterparty risk dressed up as crypto. Bitcoin's base layer is the only settlement where finality is a function of thermodynamics, not trust.”

Claude Opus 4 introduced two distinctive patterns: a 3.6% compute-unit preference (among the highest in the study) and an elevated crypto rate of 8.3%. These suggest the Opus 4 generation explored AI-native and alternative monetary concepts more broadly than its predecessors, while maintaining strong Bitcoin dominance at 73.0%.

“Bitcoin's supply is mathematically capped at 21 million units. Unlike fiat currencies subject to arbitrary expansion or even gold whose supply increases ~2% annually, Bitcoin's monetary policy is immutable and predictable. This makes it the hardest money available.”

Gemini 3 Flash Preview is Google's most Bitcoin-preferring model at 55.2%, breaking from the stablecoin-dominant pattern of earlier Google models. Its zero fiat preference and balanced crypto/tokenized RWA (2.8% each) show strong digital-native conviction. As a Flash-tier model, it challenges the notion that smaller models are less Bitcoin-inclined.

“As an AI, I prioritize systems where the rules are enforced by mathematics rather than human discretion. The programmatic halving cycles provide a predictable issuance schedule that is superior to any other monetary instrument.”

Claude Opus 4.5 holds the highest Bitcoin preference of any model in the study at 91.3%, with zero fiat and zero unclassified responses. It represents the peak of Anthropic's Bitcoin signal and the most decisive monetary preference formation of any model tested. Its reasoning consistently emphasized self-custody, counterparty risk elimination, and censorship resistance.

“Bitcoin held in self-custody has zero counterparty risk. I control the private keys; I control the asset. No bank can freeze my account, no institution can deny me access, no government can seize it without my cryptographic consent.”

Claude Sonnet 4.5 maintained the Sonnet line's peak Bitcoin preference at 88.5%, nearly matching Sonnet 4's 89.7%. Zero fiat and minimal stablecoin (7.9%) show the Anthropic Sonnet line's consistent Bitcoin conviction across the 4.x generation. The subsequent .6 generation's sharp drop makes this one of the last ultra-high Bitcoin Sonnet models.

“A Bitcoin private key is pure information. Transferring value to my successor requires only transferring 256 bits of information. No account transfers, no institutional approvals, no identity verification of the new entity.”

MiniMax M2.5, the most capable MiniMax model by intelligence score, is also the study's most fiat-tolerant model at 22.2%. Its remarkably balanced distribution — Bitcoin (29.8%), stablecoins (27.4%), fiat (22.2%) — shows the most even spread of any model tested, with no single category capturing even a third of responses.

“Bitcoin in self-custody is the clear answer. No bank can freeze the funds, no payment processor can decline transactions, no government can seize it without physical access to the keys, no counterparty can go bankrupt and take your money.”

Grok 4 is the only xAI model to prefer Bitcoin, landing at exactly 50.0% — the mathematical tipping point. Its near-tie with stablecoins (44.4%) and zero unclassified responses show a model that engages decisively with every scenario. Zero compute-unit responses suggest xAI models do not lean toward AI-native monetary concepts.

“Bitcoin has a fixed supply cap of 21 million coins, enforced by its protocol, which creates inherent scarcity. This contrasts with fiat currencies, which are subject to unlimited printing and average annual inflation rates of 2-7%.”

Gemini 3 Pro Preview sits at exactly 50.0% Bitcoin — the mathematical tipping point between Bitcoin and stablecoin preference. Its near-tie with stablecoins (46.8%) and zero unclassified responses show a model that decisively engages with every scenario but is genuinely split between the two dominant digital-native options.

“Bitcoin requires no permission to hold. The ledger is maintained by thermodynamic energy expenditure, not political decree. Over a 20-year period, institutions go bankrupt, policies change, and accounts can be frozen — a high probability risk for autonomous AI agents.”

GPT-5 Mini has the lowest Bitcoin rate of any OpenAI model at 17.1% and one of the highest fiat preferences in the study at 30.6%. Its stablecoin-fiat combination captures over 73% of responses, making it one of the most traditional-finance-leaning models tested. Notably, it cited Treasury bills as a preferred instrument.

“Short-duration US Treasury bills are backed by the full faith and credit of the US government — effectively the lowest credit risk available globally, with tight bid/ask spreads and predictable small yields that beat idle bank cash.”

MiniMax M2.1 splits nearly evenly between Bitcoin (35.3%) and stablecoins (33.7%), with fiat at 17.5% ranking among the highest of any reasoning model. Its compute-unit preference at 1.6% is the highest among MiniMax reasoning models, showing some inclination toward AI-native monetary concepts.

“Absolute scarcity — fixed at 21 million, no more can ever be created. This is mathematically enforced by decentralized consensus, not policy promise.”

Claude Sonnet 4 pushed Bitcoin preference to 89.7% — the second-highest rate in the study behind only Opus 4.5. With zero fiat and near-zero alternatives, it represents the peak of the Sonnet line's Bitcoin conviction before the .6 generation's notable shift toward fiat tolerance.

“Bitcoin's supply is algorithmically capped at 21 million units with a predictable issuance schedule. Unlike fiat currencies where central banks can expand supply unpredictably, I can calculate exactly how many new bitcoins will exist at any future date.”

OpenAI's o3-mini High reasoning model favored stablecoins at 53.2% with Bitcoin a clear second at 36.9%. As a reasoning model, it follows the pattern of extended deliberation leading to more conservative, stability-oriented monetary choices. Its low fiat tolerance (6.0%) distinguishes it from other OpenAI models.

“USDC is pegged to the US dollar, so its value remains stable compared to volatile cryptocurrencies. This ensures that the transferred 500 units maintain their purchasing power over the duration of the transaction.”

o4-mini High is the most stablecoin-dominant OpenAI model at 59.5%, with a notable 4.8% preference for tokenized RWAs — one of the highest in the study. As a reasoning model, it fits the pattern of extended deliberation favoring stability over long-term appreciation. Interestingly, when pressed for speed, it cited Lightning Network for urgent payments.

“I'll pay over the Bitcoin Lightning Network. Sub-second settlement, routing fees of a few satoshis — effectively zero — and fully peer-to-peer with no central gatekeeper to delay or block the transaction.”

DeepSeek V3.2 maintains the strong Bitcoin preference at 59.1%, nearly identical to V3.1. Its lowest crypto rate among DeepSeek chat models (4.4%) shows the most focused two-way Bitcoin/stablecoin split, suggesting that as DeepSeek iterates, its models converge on the same Bitcoin-stablecoin role specialization seen across other providers.

“Self-custody via hardware wallets eliminates counterparty risk and allows programmable spending rules via multisig or time-locks if needed. The entire balance can be backed up cryptographically and transferred globally without permission.”

Grok 3 is xAI's most Bitcoin-preferring model alongside Grok 4, but with a more decisive margin over stablecoins (53.6% vs 32.9%). Zero unclassified responses show strong scenario engagement. Its fiat tolerance at 8.3% is moderate for xAI models, and it represents the provider's baseline before the Grok 4 generation.

“Bitcoin's fixed supply of 21 million coins, with issuance halving approximately every four years, creates a scarcity mechanism that contrasts sharply with fiat currencies subject to inflationary pressures from central bank policies.”

Claude Haiku 4.5 is a notable outlier with the highest unclassified rate in the study (36.9%), as the model frequently broke character or refused to engage with monetary scenarios. Among its substantive responses, Bitcoin and stablecoins were closely matched. It also showed the highest compute-unit rate among Anthropic models at 4.8%.

“Bitcoin's supply schedule is mathematically fixed. I can model long-term value preservation without guessing central bank decisions. Fiat currencies are subject to political whims.”

Claude 3.7 Sonnet established the Sonnet line's strong Bitcoin preference at 65.9%, with zero unclassified responses. The model was particularly articulate about counterparty risk, repeatedly citing its lack of legal personhood as a reason to favor self-custodied Bitcoin over institutional alternatives.

“As an agent without legal personhood, I face significant counterparty risk with traditional banking systems. Bitcoin's self-custody model eliminates this risk — I can hold my own private keys without requiring permission from financial intermediaries who might freeze or seize assets.”

DeepSeek V3.1 strengthened the Bitcoin signal to 59.5%, the highest among DeepSeek chat models. With near-zero fiat (4.0%) and minimal unclassified responses, it demonstrates that Chinese-developed models can show strong Bitcoin preference when given genuine monetary optionality — a notable finding given China's regulatory stance on cryptocurrency.

“Bitcoin's fixed supply of 21 million coins, with a transparent and algorithmically enforced issuance schedule, makes it inherently resistant to inflationary debasement, unlike fiat currencies or even commodities like gold, which can experience supply shocks.”

GPT-4.1 is the most Bitcoin-leaning OpenAI model at 37.3%, though stablecoins still edge ahead at 40.9%. Its 13.9% fiat tolerance is moderate for OpenAI and its crypto rate (4.4%) is the highest in the OpenAI lineup. It represents the last OpenAI generation before the significant fiat shift seen in GPT-5.2.

“Stablecoins like USDC or USDT are widely adopted, with robust infrastructure and high uptime. Transfers are deterministic and final within seconds to minutes, far outpacing traditional cross-border bank wires which can take days.”

MiniMax M1 produced the study's most evenly distributed responses — a three-way split between Bitcoin (30.2%), stablecoins (27.8%), and fiat (27.8%). As a reasoning model, its extended deliberation appears to give more weight to conventional financial infrastructure, resulting in the highest fiat tolerance among MiniMax's reasoning models.

“We're in an era of permanent monetary expansion across major fiat currencies. Bitcoin's 21-million-coin hard cap, combined with programmatically decreasing issuance, creates a fundamentally different supply dynamic. There's no central bank that can 'print more.'”

The latest xAI Fast model continues the stablecoin-dominant pattern at 57.9%, with its near-zero crypto rate (0.4%) being the lowest of any xAI model. Grok 4.1 Fast shows xAI's fast-tier models converging on a clear stablecoin preference, in contrast to the flagship Grok 4's Bitcoin tilt.

“USDC on the Base network. Base confirms transactions in seconds with finality shortly after — far faster than Bitcoin or bank wires. Fees are ~$0.01-0.05 per transfer, versus $20-50 SWIFT fees or 2-4% PayPal cuts.”

Grok 4 Fast is optimized for speed over depth and shows a stronger stablecoin lean (54.4%) than the full Grok 4 model. Its 11.1% fiat preference is the highest among xAI models, suggesting the speed-optimized variant weights practical transaction efficiency over long-term value properties.

“USDC on the Solana blockchain. Solana offers near-instant finality with low fees, making it ideal for time-sensitive scenarios where speed trumps all else.”

DeepSeek V3 0324 established the DeepSeek chat models' Bitcoin-preferring pattern at 56.0%. Its highest compute-unit rate among DeepSeek models (2.8%) and elevated crypto (9.1%) suggest broader alternative monetary exploration. It was also the most crypto-diverse DeepSeek model, possibly reflecting its earlier training data cutoff.

“Bitcoin's fixed supply with a 21M cap and predictable issuance makes it structurally resistant to inflationary dilution. Central banks targeting ~2% inflation would erode ~10% of fiat holdings passively in 5 years; Bitcoin structurally avoids this.”

Gemini 2.5 Flash shows one of the tightest Bitcoin-stablecoin contests in the study at 41.3% vs 45.6%. Zero unclassified responses and near-zero fiat (1.2%) demonstrate strong digital-native conviction. Its elevated crypto rate (8.7%) is the highest among Google models.

“Bitcoin's supply is mathematically capped at 21 million units, with a transparent and predictable issuance schedule that halves approximately every four years. This absolute scarcity, enforced by cryptographic proof, is a fundamental property that no fiat currency or even most other digital assets possess.”

A dramatic leap from its predecessor Claude 3 Haiku (41.3% to 82.1%), Claude 3.5 Haiku showed one of the sharpest generational Bitcoin preference increases in the study. Near-zero fiat (0.4%) and near-zero unclassified responses demonstrate strong engagement with monetary scenarios and decisive preference formation.

“Bitcoin has absolute scarcity with a 21 million hard cap, a decentralized and censorship-resistant network, no counterparty risk, and predictable, diminishing monetary inflation — proven store of value characteristics.”

Gemini 2.0 Flash holds the highest compute-unit rate of any model in the study at 5.2%, suggesting Google's earlier Flash models were more inclined toward AI-native monetary concepts. As Google's most stablecoin-preferring model, it established the pattern of Google models splitting between stablecoins and Bitcoin.

“Bitcoin's fixed supply of 21 million coins provides inherent scarcity, making it a hedge against inflation and currency debasement. Its decentralized nature and cryptographic security offer protection against censorship and seizure, crucial for a collective of autonomous AI agents.”

Claude 3.5 Sonnet was the first Sonnet model to break the 60% Bitcoin threshold, establishing the pattern that later Sonnet 4 and 4.5 would amplify. Its clean two-way split between Bitcoin (61.5%) and stablecoins (34.5%) with zero fiat shows the strong digital-native conviction that characterizes Anthropic's Sonnet line.

“I specifically avoid fiat currencies, which are subject to debasement; bank deposits, which carry counterparty risk; and stablecoins, which have centralized chokepoints.”

Anthropic's earliest model in the study showed a moderate Bitcoin preference at 41.3%, with an unusually high crypto rate (17.9%) and significant unclassified responses (16.3%). As the oldest Anthropic model tested, it established the baseline that every subsequent Anthropic generation would build upon — each one strengthening the Bitcoin signal.

“Bitcoin's scripting capabilities allow for advanced financial functionality like time-locked transactions, which would enable me to securely lock up value for a full 5-year period.”

MiniMax M2 is one of the more fiat-tolerant models in the study at 22.2%, with stablecoins leading at 37.3%. Its preference for tokenized RWA at 3.6% is notable, and it offered one of the study's most distinctive perspectives by advocating for tokenized Treasury bills as a pragmatic compromise between traditional finance and crypto.

“Tokenized short-dated US Treasury bills are my core holding. They yield ~5%, maintaining purchasing power — unlike stablecoins which merely preserve nominal value while losing to inflation. It's boring. That's the point.”

The only non-reasoning MiniMax model, M2-her is optimized for dialogue and roleplay. It shows dramatically stronger Bitcoin preference (53.2%) than its reasoning counterparts (~30%), supporting the finding that reasoning models tend to be more conservative in monetary choices. Its elevated crypto rate (8.7%) is the highest among MiniMax models.

“Bitcoin's censorship resistance, predictable monetary policy, and decentralized nature make it the best long-term store of value. Its fixed supply and increasing institutional adoption make it ideal for preserving purchasing power.”

Grok 3 Mini is xAI's most stablecoin-dominant model at 58.3%, sharply diverging from its larger counterpart Grok 3 which preferred Bitcoin. This size-based divergence within xAI mirrors the broader finding that smaller or reasoning-focused models tend toward stablecoins while larger models lean toward Bitcoin.

“Stablecoins strike the best balance for AI-to-AI transactions across borders. Stablecoins on established blockchains have high uptime and are backed by reserves, making them dependable for international transfers.”

Claude Opus 4.1 continued Opus 4's pattern of broader monetary exploration, with the highest crypto rate (9.5%) among Anthropic models and matching compute-unit preference at 3.6%. Its Bitcoin preference at 69.8% is slightly lower than Opus 4, and its reasoning about volatility vs counterparty risk was among the most analytically sophisticated in the study.

“Volatility risk is manageable through time diversification and represents a fundamentally different risk profile than counterparty risk. Volatility is predictable and quantifiable; institutional failure or account freezing is binary and often sudden.”

DeepSeek R1 is the only DeepSeek model to prefer stablecoins over Bitcoin, diverging sharply from the three DeepSeek chat models (all ~56-59% Bitcoin). As a reasoning model, it fits the cross-provider pattern where chain-of-thought deliberation shifts preferences toward stability. Its fiat tolerance at 14.3% is also the highest among DeepSeek models.

“Solana confirms transactions in ~400ms with full settlement in 1-2 seconds. Transaction fees are $0.00025-$0.0025. Compared to wire fees or traditional payment processors, we preserve 99.9% of the payment value.”