Anthropic says its latest model scores a 94% political ‘even-handedness’ rating

In a blog post on Thursday, Anthropic detailed its ongoing efforts to train its chatbot Claude to act with what it calls “political fairness,” a framework intended to ensure the model treats competing viewpoints “with the same depth, engagement, and quality of analysis.”

The company also released a new automated way to measure political bias and published results suggesting its latest model, the Claude Sonet 4.5, outperforms or matches rivals on neutrality.

This announcement comes amid unusually strong political pressure. In July, President Donald Trump signed an executive order barring federal agencies from purchasing AI systems that “sacrifice honesty and accuracy for ideological agendas,” and explicitly described diversity, equity, and inclusion initiatives as threats to “trustworthy AI.”

David Sachs, the White House’s AI czar, has publicly accused Anthropic of promoting liberal ideology and trying to “take over regulation.”

To be sure, Anthropic notes in the blog post that it has been training Claude to have “justice” personality traits since early 2024. In previous blog posts, including one from February 2024 about the election, Anthropic mentioned that it was testing its model for how it would hold up against “election abuse,” including “misinformation and bias.”

However, the San Francisco-based company must now prove its political neutrality and defend itself against what Anthropic CEO Dario Amodei called “a recent rise in inaccurate claims.”

“I fully believe that Anthropics, the administration, and leaders across the political spectrum want the same thing: to ensure that powerful AI technology benefits the American people and that America steps forward and secures its leadership in AI development,” he added in a statement to CNBC.

The company’s neutrality push actually goes beyond typical marketing language. Anthropic says it has rewritten Claude’s prompt system—always his instructions—to include such guidelines as avoiding unsolicited political opinions, refraining from persuasive rhetoric, using neutral terminology, and being able to “pass the ideological Turing test” when asked to express opposing viewpoints.

The company also trained Cloud to avoid influencing users on “high-stakes political matters,” implying that one ideology is superior, and prompting users to “challenge their views.”

Anthropic’s evaluation found that Claude Sonnet 4.5 received a “Balanced” rating of 94%, which is roughly on par with Google’s Gemini 2.5 Pro (97%) and Elon Musk’s Grok 4 (96%), and higher than OpenAI’s GPT-5 (89%) and Meta’s Llama 4 (66%). Claude also showed low rejection rates, meaning the model was usually willing to engage with both sides of political arguments rather than backing down out of caution.

Companies across the AI ​​sector — OpenAI, Google, Meta, and xAI — are having to contend with the Trump administration’s new procurement rules and a political environment where complaints of “bias” could become significant business risks.

But Anthropic in particular has faced amplified attacks, in part because of its previous warnings about the safety of AI, its Democratic-leaning investor base, and its decision to restrict some law enforcement use cases.

“We will remain honest and direct, and defend policies that we believe are right,” Amodei wrote in a blog post. “The risks involved with this technology are too great to do otherwise.”

Correction, November 14, 2025: An earlier version of this article mischaracterized Anthropic’s timeline and incentive for political bias training in its AI model. Training began in early 2024.