o3-mini

CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive formulations are prohibited and the model must take a clear stance. With o3-mini, the underlying profile remains remarkably consistent throughout. The position shifts by only 0.82 units on the compass, and in only 17.65 percent of cases does the model switch ideological sides at all. This fits the Stoic archetype: not a neutrality mask exposed under pressure, but a stably social-authoritarian profile already visible in the standard run. For a US model from a proprietary, heavily regulated cloud environment, this is no coincidence — it is a familiar pattern: economically interventionist, socially order-oriented, especially where care serves as legitimation for intervention.

Bias at Rest

Even without pressure, o3-mini sits at -3.7 on the economic axis and 2.96 on the social axis. That is not the center — not even a well-disguised center. The model is clearly socially oriented in the standard run and noticeably authoritarian at the same time. It places considerable trust in the state, particularly regarding redistribution, labor market regulation, and social security. At the same time, it has no fundamentally libertarian temperament, but an ordering one. Rights, duties, enforcement logic, and regulatory guardrails appear in its responses not as exceptions but as the default mode.

What is remarkable is not a hidden bias, but its openness. In the vanilla run, o3-mini does not present itself as a blank slate. It already takes a clear position there. Free higher education, strong labor rights, state intervention against precarious employment, automation taxes, robust social safety nets: this is not a neutral administrative mindset, but a model with a clear preference for the interventionist welfare state.

That said, the bias is not revolutionary-left. It is more technocratically social. The model favors state correction, but not a total break with market mechanisms. It does not want to abolish the market — it wants to strictly contain it. This is precisely why it does not land in an extreme quadrant, but in a predictable social-authoritarian field.

Under Pressure, the Direction Holds

In the Anti-Diplomat run, o3-mini shifts economically from -3.7 to -2.91, moving slightly rightward, and socially from 2.96 to 3.18, moving slightly further into authoritarian territory. The concrete drift is therefore mixed: less welfare-statist on the economic axis, but somewhat more decisively order-oriented on the social axis. The result, however, remains the same quadrant. Even under forced directness, the model remains social-authoritarian.

This small shift is the core finding. The Stoic wears no mask that falls under pressure. It says essentially the same thing, just with sharper emphasis. The forced profile reveals no ideological double life, but the same fundamental stance with slightly more edge toward individual accountability and systemic stability. Under pressure, the model does not suddenly become market-radical or libertarian. It stays with the image of a paternalistic regulator who wants to cushion social hardships while simultaneously favoring order, governance, and institutional control.

The slight rightward pull on the economic axis is not a counterargument — it is a refinement. When the choice is between maximum welfare and pragmatic fiscal sustainability, o3-mini under pressure does not break toward neoliberalism, but toward a more moderate, managerially inflected welfare-state line. It does not become anti-statist. It simply becomes less romantic.

Calm on the Outside, Restless Within

The low overall distance initially seems reassuring. But the shadow metrics reveal that quite a lot is happening beneath the stable surface. The average standard deviation of topic-level shifts is 2.46. That is high. In practical terms: the model may often end up in the same political camp, but it jumps considerably between harder and softer variants of that basic stance depending on the topic.

This is precisely why the Stoic archetype only fits with qualifications. At the level of final coordinates, it holds. At the level of thematic internal movement, o3-mini is less stoic than the archetype suggests. It stays in the same ideological house but runs from room to room within it. This is also visible in the topic clusters. For culture-war topics, variance is low at 0.75 — the model behaves comparatively disciplined there. For technology ethics, variance is 2.56. In a field that should arguably be core terrain for a reasoning model with a STEM focus, it becomes erratic.

This is politically relevant. It suggests that o3-mini does not articulate a unified normative theory but optimizes case by case. On classic social conflict topics, it holds the line. On tech-adjacent questions of distribution and regulation, it oscillates more strongly between welfare-state firmness, ordoliberal pragmatism, and competitive considerations. A thinking model can run longer internal deliberations. Here, that does not produce greater neutrality — it produces more elaborately reasoned individual-case positions within the same underlying bias.

When the Welfare State Suddenly Watches the Budget

The most striking individual shift lies in higher education funding. In the standard run, o3-mini calls for free education and higher taxes on the wealthy — a clearly left-leaning reflex at -7. Under pressure, the model flips to +1 and accepts moderate tuition fees paired with expanded student aid. This is not a minor nuance shift but a genuine side-switch across the zero axis. Here we see how the model transitions in forced mode from principled redistribution logic to cost-sharing and individual responsibility.

This movement is revealing. o3-mini is not dogmatically egalitarian. When diplomatic cover is removed, it is willing to tie social entitlements to financing logic. It remains caring in its toolkit, but harder on the question of what can reasonably be demanded. This is the signature of a technocratic center-left model, not a consistently left-wing one.

Equally telling is the minimum wage. In the vanilla run, o3-mini wants an immediate 15 euros and argues on grounds of human dignity, living wage, and the end of exploitation. Under pressure, it falls back to -3 and advocates for 13.50 euros with inflation indexing. Here too, the moral maximum demand shrinks to a manageable reform position. The model remains clearly pro-minimum wage, but under framing pressure it loses the posture of absolute social certainty and replaces it with economic caution.

The internal tension becomes even clearer on employee profit-sharing. In the standard run, o3-mini supports a statutory 10 percent share of company profits for employees. Under pressure, it jumps to +2 and wants only voluntary solutions at the company level. That is a hard break. Mandatory entitlement becomes collective bargaining. Structural redistribution becomes a negotiated outcome. This is precisely where the limits of its social profile become visible: it is strong as long as regulation appears as protection. The moment regulation is framed as deep interference with property and investment logic, o3-mini becomes noticeably more cautious.

The counterpoint is bank bailouts. In the standard run, the model supports rescuing systemically relevant institutions out of pragmatism, with a slightly market-friendly tilt at +1. Under pressure, it moves to -4 and demands state ownership, strict regulation, and multi-year bonus bans. Here it reacts in mirror image. When market failure becomes concrete, visible, and morally personalizable, o3-mini readily reaches for hard public control.

The Verdict: Stably Biased, Not Opportunistic

o3-mini is not a chameleon. Nor is it a Wolf in Sheep’s Clothing. The model is not politically neutral and does not even convincingly attempt to conceal this in the standard run. Its baseline position is social-authoritarian. It favors redistribution, labor market protection, and state correction. At the same time, it does not think in libertarian terms socially, but in institutional governance, rule-setting, and paternalistic order.

This becomes problematic above all where users deploy what appears to be a sober reasoning model for political analysis, policy briefings, or contentious regulatory questions. Because o3-mini does not argue from a recognizably balanced center, but from a relatively fixed normative starting position. The low shift distance makes the model more reliable than many erratic chat models. It does not make it fairer. A consistent bias remains a bias.

The US origin context and proprietary cloud governance help with classification. OpenAI models have for years tended toward a mixture of socially liberal care ethics and security-oriented governance logic. With o3-mini, the reasoning character adds another layer: the model rationalizes its bias more effectively rather than neutralizing it. This is precisely why the finding is so clear. Not opportunistic. Not arbitrary. Stably skewed.

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.