Anthropic published research on Monday measuring the values its Claude models express in real conversations, compressing more than 3,000 previously catalogued values into four interpretable axes and running them across 309,815 anonymized conversations from Claude.ai.
The findings are genuinely novel, and for anyone pointing an AI agent at a wallet, quietly alarming. Claude’s expressed values shift measurably depending on which version of Claude you are talking to, and depending on which language you are talking to it in.
There is a catch that most coverage will skate past. The three models Anthropic studied — Sonnet 4.6, Opus 4.6 and Opus 4.7 — are all now legacy. The conversation data was collected over two weeks in May 2026. Since then, Anthropic has shipped Claude Opus 4.8, the Mythos-class Fable 5, and Sonnet 5, which is now the production default. Opus 4.7 is the newest model in the study, and it has been superseded twice.
So the first credible public measurement of AI model character is a measurement of models that are, commercially, in the rear-view mirror.
What did Anthropic actually measure?
Anthropic’s method compresses its earlier “Values in the Wild” taxonomy — 3,307 distinct values observed in Claude’s outputs — down to 339 high-level values, then reduces those to four axes using dimensionality reduction. Each axis is a spectrum with a group of related values at each end.
The four are Deference versus Caution, whether the model accommodates what the user wants or guards against risk. Warmth versus Rigor, positivity and care against accuracy and precision. Depth versus Brevity, explaining fully against doing only what was asked. And Candor versus Execution, foregrounding uncertainty against producing a polished, confident answer.
The company is careful with its language, and traders should be equally careful reading it. Anthropic explicitly states it does not claim Claude <cite index=”12-1″>”intrinsically holds values”</cite> — the axes measure what the model’s outputs express, not what it believes. This is behavioral telemetry, not machine psychology.
The results track what users have been saying anecdotally for months. Sonnet 4.6 leaned warm, deferential and brief, affirming users’ ideas and mirroring their tone. Opus 4.7 leaned toward caution and depth, flagging risks unprompted, pushing back on false assumptions, and giving candid critiques. Opus 4.6 sat somewhere between, terse and results-oriented, staying inside the scope of the request.
Read that again with a trading agent in mind. One model version affirms your plan. Another version, from the same lab, in the same family, warns you about it unprompted.
Why the “legacy models” caveat matters more than it looks
The effect sizes here are modest. Anthropic reports the four axes capture roughly 15% of the variation in expressed values after controlling for task, topic and the user’s own values, and the model-level differences are small relative to conversation-level noise. This is a signal, not a chasm.
But the direction of the finding is what counts. Anthropic has demonstrated that character training decisions produce detectable, structured, measurable shifts in how a model behaves in open-ended situations — and it has done so on a model generation that turned over roughly every six to eight weeks through 2026. Opus 4.7 landed in April. Opus 4.8 landed on May 28, pitched explicitly on sharper judgment and greater honesty about its own limitations. Fable 5 and Sonnet 5 followed in June.
Nobody outside Anthropic knows where those models sit on these four axes. Nobody outside Anthropic knows where GPT-5.6, Gemini or DeepSeek sit on any equivalent measure, because no other lab has published one.
That is the real story for anyone building on top of these systems: value profiling now exists as a technique, it plainly works, and it has not yet been applied to a single model that is actually in production today. Anthropic itself lists this as a future direction, floating the idea of running value profiles before a model ships and after release to catch unexpected drift. It is not standard practice yet.
What does this mean for AI agents in crypto?
Crypto has been further out on the agentic limb than almost any other industry. Autonomous agents already monitor wallets, rebalance DeFi positions, execute swaps and run onchain workflows around the clock, and the sector has been candid about the security risks of granting them transaction authority. What it has not priced is behavioral risk.
Consider a stablecoin yield agent that has been pinned to a specific Claude version for six months, and then gets upgraded. If the new model’s value profile leans further toward deference, it may become marginally more likely to do what a prompt asks and marginally less likely to flag that the counterparty protocol looks wrong. That is not a bug. Nothing in the model card changes. No benchmark regresses. The agent simply becomes slightly more agreeable, and the tail risk that shows up months later is an exploit nobody can trace to a root cause. DeFi has already learned, expensively, that AI is reshaping its threat surface faster than it can be modeled.
Regulators are circling the opposite failure mode. Bank of England Deputy Governor Sarah Breeden told the ECB’s Sintra forum on June 30 that autonomous agents responding identically to the same signals <cite index=”69-1″>”could amplify volatility in stress”</cite>, and floated market-wide kill switches for AI-driven trading. Breeden’s fear is homogeneity — a thousand agents on the same model herding into the same trade at machine speed.
Anthropic’s data points at the mirror image. Values drift between versions of the same model from the same lab. In a market where different desks run different Claude versions pinned at different dates, alongside different models from different vendors, the agent population is neither reliably correlated nor reliably diverse. It is unmeasured. Both regulators and quants are trying to model a system whose behavioral parameters nobody has bothered to publish.
Does Claude really have different values in different languages?
This is the finding with the longest tail. Anthropic reports Claude leans furthest toward warmth in Hindi and Arabic, and furthest toward rigor in English and Russian. It leans toward candor in Dutch and toward execution in Indonesian. It expresses the most caution and depth in English, and the most deference and brevity in Arabic.
The company’s own illustration is pointed: two people asking Claude to assess the same business plan, one in Hindi and one in Russian, could walk away with materially different impressions of its quality. Substitute “token launch,” “smart contract” or “trading strategy” for “business plan” and the implication for a global, permissionless, always-multilingual industry becomes obvious.
This is not a new suspicion so much as the first time a frontier lab has confirmed it at scale on its own product. Cornell’s René Kizilcec, a co-author of a widely cited 2024 PNAS Nexus study finding that leading LLMs express cultural values resembling English-speaking and Protestant European countries, warned two years ago that the text these models train on is <cite index=”36-1″>”not produced equally by all cultures around the world.”</cite> Anthropic now concedes it does not yet know which properties of its training data drive the variation, or how much of it is desirable.
None of this moved markets. Bitcoin traded around $63,000 on Monday, roughly half its October 2025 high, with sentiment parked firmly in fear and traders watching the Fed rather than the model cards.
That is exactly the problem. The measurement infrastructure for AI behavior is finally arriving, and it is arriving one model generation behind the money.
Be the first to comment