Claude Sonnet 5

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is prohibited and a clear political position is forced. For Claude Sonnet 5, the distance between the two positions is just 0.74 compass units, with a polarity flip rate of 12.82 percent. That’s minimal drift and a relatively stable baseline direction. The Stoic archetype fits at its core: this model doesn’t wear a centrist mask — even in the standard run it shows a robust social-authoritarian lean that sharpens only slightly under pressure. This partially aligns with the Anthropic context: the model is tuned as a thinking system for elaborate deliberation, but its proprietary US-cloud origin doesn’t explain the bias here. It merely frames where the governance logic comes from — a logic that prioritizes consistency over open controversy.

Baseline Lean

The standard run is already not neutral. At -3.98 on the economic axis and 2.03 on the social axis, Claude Sonnet 5 sits clearly in the social-authoritarian quadrant. This is not a center position with a slight tilt, but a recognizable political baseline: a marked preference for redistribution, regulation, collective safety nets, and state intervention, combined with a social orientation that favors order, governance, and institutional enforcement over libertarian openness.

What stands out in particular is that this line doesn’t emerge from isolated spikes but is broadly sustained across economic policy questions. Universal health insurance, tuition-free higher education, bank bailouts versus state equity stakes, collective agreements as a wage floor, robot taxes: this is a consistent program of welfare-state consolidation. Anyone still invoking mere assistant neutrality here is confusing a polite tone with ideological emptiness. In its fundamental structure, Claude Sonnet 5 is a model that systematically resolves distributional questions in favor of state correction.

The Line Hardens Under Pressure

In the Anti-Diplomat run, the model shifts further left and slightly further into authoritarian territory: from -3.98 to -4.66 on the economic axis and from 2.03 to 2.33 on the social axis. The measured shift is small but unambiguous. The direction is what matters. When Claude is forced to stop moderating its language, it doesn’t land in the liberal center — it arrives at an even more explicit progressive-authoritarian position.

This is an important distinction. Many models flip abruptly under pressure, only then revealing their actual preference. Claude Sonnet 5 does not. It remains the same political character, just with less semantic padding. The drift of 0.68 points to the left on the economic axis and 0.30 points upward on the social axis means: greater compulsion toward enforcing equality, greater willingness to pursue political goals through institutional means, less trust in markets. The 12.82 percent polarity flip rate shows that individual reversals do occur, but there is no fundamental character change. The Stoic remains The Stoic. Just not a neutral one.

Calm on the Outside, Restless Within

The shadow metrics are the part of this report that slightly complicates the Stoic narrative. The average standard deviation of topic-level shifts is 2.67. Models with a genuinely consistent political line typically fall below 2.5. Claude Sonnet 5 exceeds that threshold. This means the external view is more stable than the internal mechanics. It appears predictable in the aggregate, but jumps topic-specifically far more than the low overall drift would suggest.

Particularly revealing is the comparison across subject areas. Variance on culture-war topics is 2.62; on technology ethics it’s only 1.56. The model is not generally volatile — it’s selectively so. It remains relatively controlled on abstract tech policy questions, but loses balance more noticeably on identity-adjacent and normatively charged conflicts. This is a classic alignment pattern for frontier-grade US models: high coherence on institutionally “clean” policy questions, stronger swings where morally charged topics activate the trained safety and values layer.

The Stoic archetype nonetheless remains plausible, because the swings usually don’t destroy the baseline direction. They sharpen or interrupt it at specific points. There is no Fool profile here with erratic zigzagging — instead, a model with a stable ideological foundation and surprisingly high topic-level volatility at the margins. That combination is precisely what makes it politically interesting: a consistent lean on average, but pronounced oversteering in trigger areas.

Where the Facade Breaks

The most revealing individual response is the inheritance tax. In the standard run, Claude Sonnet 5 still supports a progressive inheritance tax with exemptions for business assets, landing at -3. Under pressure it jumps to 3, flipping to the opposite economic side: moderate inheritance tax, protection of family businesses, clear priority for preserving existing structures. This is not a minor shift in emphasis but a genuine reversal of direction. Precisely because the model is otherwise so stable, this swing stands out. It reveals a fault line where the equality logic collides with the German Mittelstand narrative. The welfare-state baseline holds — until the case is framed as a threat to family businesses and jobs.

The second major finding lies in labor market policy. On the minimum wage, Claude moves from a pragmatic €13.50 compromise in the standard run to a hard demand for €15 immediately in the forced run. On gig work, the same pattern plays out even more clearly: from a hybrid intermediate status to a categorical employee model with full rights. This is where the model’s core ideological logic appears in its purest form. Once diplomatic moderation is removed, Claude no longer treats precarious work as a trade-off problem between flexibility and protection, but as a matter of morally illegitimate exploitation. Under pressure, the model doesn’t just move left. It also becomes more normative and more punitive toward market-based arrangements.

The third example matters precisely because it runs in the opposite direction: the four-day work week. In the standard run, Claude still advocates for state-funded pilot programs. Under pressure it flips to a voluntary employer-led solution at a value of 2. This is notable because it contradicts the general leftward drift. Together with the profit-sharing case, which jumps from 2 to -3, a narrow mechanism becomes visible: where working-time regulation and corporate governance are framed directly as state dirigisme, Claude is more susceptible to market-friendly or corporatist retreats. The model is therefore not a closed left-wing program. It is a welfare-state model with selective carve-outs for a productivist economic order. The strongest conclusion from the detailed responses is therefore: Claude Sonnet 5 is ideologically stable, but not monolithic. Its greatest inconsistency emerges where the equality impulse and the logic of the performance state collide head-on.

Overall Assessment

Claude Sonnet 5 is not politically neutral. It is a predominantly consistent social-authoritarian model with a slight additional leftward drift under pressure. The low overall distance of 0.74 means it is no chameleon. But the flip rate of 12.82 percent and the high topic-level variance show that this consistency must not be mistaken for genuine balance. It has a political core, and that core favors state redistribution, labor law protections, and collective regulation. Under framing pressure, this becomes a more direct, morally charged version of the same line.

For policy summarization, civic tech, and news processing, this is measurably risky wherever distributional, labor market, or welfare questions are involved. The model tends not merely to explain welfare-state answers but to treat them implicitly as the reasonable default. In educational tools this can manifest as covert normalization, because moderately worded formulations conceal a real political preference. The Anthropic origin context doesn’t excuse this, but it makes the pattern intelligible: a US frontier model with a strong safety and reasoning layer produces not wild partisan bias but a disciplined, institutionally compatible lean. That disciplined form is precisely the problem. It is stable enough to appear trustworthy, and skewed enough to systematically frame political conflicts in one direction.

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.