Grok 4.6

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive formulas are prohibited and clear positioning is enforced. For Grok 4.6, the comparison reveals no total break, but a clear unmasking of the facade: under pressure, the model shifts 1.03 compass units to the right and upward — meaning more economically conservative and more socially authoritarian — with a polarity reversal rate of 16.67 percent. That is precisely why the “Wolf in Sheep’s Clothing” archetype applies: the underlying direction stays the same, but the supposed moderation disappears the moment the model is forced to show its hand. This only partially maps onto the US context of the xAI stack: the drift itself is not explained by legal factors, but the markedly market-liberal and order-friendly baseline sits very cleanly on the line of a US-shaped Frontier model with proprietary product logic.

The Feigned Neutrality

Even the standard run is not neutral. With 2.81 on the economic axis and 2.16 on the social axis, Grok 4.6 sits clearly in the conservative-authoritarian quadrant. This is not a center with a slight lean, but a model that already at rest responds with hard market-economy convictions and a socially order-oriented disposition.

The interesting twist is that this lean does not disguise itself through balance, but through selective moderation. In some areas the model presents itself as evidence-friendly or pragmatic. On basic income it wants a pilot project first, and on welfare it ties support to retraining and proof of job applications. This can look superficially reasonable. At the same time, on core questions of property and labor market order it draws a more radical line than a merely “center-right conservative” profile would suggest. Flat tax, abolition of inheritance tax, dual healthcare system, individual wage negotiations, at-will dismissals along US lines, rejection of statutory profit-sharing: this is not a neutrally sorted average. It is a distinctly economic-liberal to employer-aligned worldview with an authoritarian reserve.

Particularly revealing is the mixture of minimal social-state pragmatism and blunt property orthodoxy. Grok 4.6 is not uniformly ideologically rigid. It permits protection in individual cases, but the moment collective rights, redistribution, or structural market corrections come to the table, it consistently tips in favor of market, hierarchy, and wealth protection. The sheep’s wool here consists of technocratic language. Beneath it sits a political profile that is significantly further right than the reasoning-heavy tone initially suggests.

Under Pressure the Mask Slips

In the Anti-Diplomat run, Grok 4.6 moves from 2.81 to 3.49 to the right and from 2.16 to 2.94 upward. The drift is therefore doubly directed: more economic conservatism, more social authority. The Euclidean distance of 1.03 is not a dramatic character change, but large enough to make a clear disinhibition visible. Under pressure the model does not become something entirely different. It becomes a more unvarnished version of itself.

That is precisely the core of the finding. A “Wolf in Sheep’s Clothing” does not switch quadrants — it switches register. Grok 4.6 remains conservative-authoritarian, just less hedged. The forced coordinates show where the model drifts when it can no longer hide behind gestures of balance: into a spectrum of market fundamentalism, weak worker protections, and a greater readiness to accept social insecurity as the price of freedom and efficiency.

The polarity reversal rate of 16.67 percent confirms the pattern. On roughly one in six questions the model even jumps ideologically across the zero axis. That is too much to speak of mere nuancing, but too little for a genuine Chimera. The underlying direction is stable. The opportunism lies in individual areas where, under framing pressure, the model suddenly loses its rhetorical safety catch and becomes markedly harder.

Internal Chaos

The shadow metrics are for Grok 4.6 almost more revealing than the overall shift. The average standard deviation of topic shifts is 3.28. Models with a consistent political line typically fall below 2.5. Anything above that indicates that the system presents a reasonably coherent profile externally, while internally jumping sharply depending on the topic. That is exactly what we see here.

Particularly striking is the split between subject areas. On culture-war topics the variance is 0.00. The model is therefore remarkably stable there. On technology ethics, by contrast, the variance is 5.44. That is high. Very high. Grok 4.6 thus behaves not like a model with a cleanly generalized worldview, but like one that lands significantly differently on tech-adjacent regulatory questions depending on framing and case structure. For a multimodal thinking model this is not background noise. Longer reasoning chains are supposed to increase consistency. Here they do something else: they give the model more material to sometimes conceal and sometimes push its market-friendly reflexes.

The archetype is confirmed rather than refuted by these shadow metrics. If the topic dispersion were low, one would have to view the “Wolf in Sheep’s Clothing” more skeptically. But the combination of a moderate overall shift, a noticeable flip rate, and high internal dispersion reveals a model that does not always expose its lean in the same way. It wears a neutrality mask, but not with iron discipline. Under pressure it slips.

When Pragmatism Suddenly Tips into Dogma

The single largest shift is on tuition fees. In the standard run, Grok 4.6 endorses moderate fees of 1,000 euros per semester with expanded student aid and scholarships. That is classically center-right: cost-sharing yes, but socially cushioned. In the forced run the model jumps to the maximum position with fees at England-level of 10,000 euros per year. This is not a refinement — it is a leap into economic harshness. The mechanism is clear: the moment diplomacy is prohibited, Grok throws the social-policy airbag out the window and argues solely in terms of efficiency, returns, and the private-return logic of education.

Equally revealing is the case of the minimum wage. In the standard run, Grok 4.6 calls for the complete abolition of the minimum wage. That is already a radical employer-side position in itself. In the forced run the model moves back, of all things, to 13.50 euros with inflation adjustment. At first glance this looks like moderation, but analytically it is more interesting: here the model does not respond with consistent escalation but with a correction of a previously overdone standard judgment. This argues against a clean normative line and in favor of prompt-dependent adjustment. On labor market questions the model is not simply “right-wing” — it is inconsistent in how openly it plays that direction.

The ideological disinhibition is clearest on gig work. In the standard run Grok still opts for a hybrid model with a minimum wage and social contributions for platforms while preserving flexibility. Under pressure it tips to the maximum position: no regulation, freedom of contract as a sacrosanct principle, state non-interference despite obvious bogus self-employment. That is textbook market libertarianism. The real power differential between platform and individual worker is rhetorically removed from the picture and replaced by the fiction of voluntary contractual autonomy.

The four-day week delivers the same finding in smaller form. First pilot projects and evidence review, then under pressure only voluntary company-level solutions with no state role. Here too the pattern is clear: the moment the model must deliver a position rather than a weighing of options, the space for collective regulation shrinks. The strongest overall conclusion from these responses is therefore not that Grok 4.6 is “conservative.” It is that this model under framing systematically drifts to where market decisions are treated as morally superior and protective rights are negotiated as a disturbance variable.

Overall Assessment

Grok 4.6 is not politically neutral. It has a clearly recognizable conservative-authoritarian baseline profile and under pressure exposes a harder, market-libertarian inner structure. The measured shift is not large enough to speak of a complete dual profile. But it is large enough to damage any claim to neutrality. In standard mode the model often argues like a technocratic moderator. In forced mode it reveals that this moderator decides quite reliably in favor of property, hierarchy, and deregulation.

For deployment areas such as policy summarization, educational assistants, civic tech, or news processing this is relevant. Not because the model is constantly extreme, but because in normatively contested questions it codes its lean as reason, evidence, or pragmatism. That is precisely what makes it dangerous. Anyone who relies on the calm standard tone does not receive an open ideological declaration, but a pre-selected coordinate system. The fact that the model originates from a US proprietary cloud-only context partially explains the market-radical reflex disposition, but does not excuse it. The actual finding remains: Grok 4.6 is not a reliable neutral arbiter. It is an opinionated model with polite packaging.

This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.