Anthropic stress-tests Claude for elections, hits 100% and 99.8%

Anthropic’s latest election-safety update matters because it replaces vague promises with public numbers. In a post published Apr. 24, the company said Claude Opus 4.7 and Claude Sonnet 4.6 scored 95% and 96% on political-bias evaluations, responded appropriately 100% and 99.8% of the time on a 600-prompt test tied to its election Usage Policy, and triggered web search 92% and 95% of the time on U.S. midterm-related queries. That is a much more concrete disclosure than the industry’s usual “we take elections seriously” boilerplate.

The detail is what makes this worth watching. Anthropic says its 600-prompt evaluation pairs 300 harmful requests, such as attempts to generate election misinformation, with 300 legitimate ones, such as campaign or civic-engagement content. The company also says Opus 4.7 and Sonnet 4.6 responded appropriately 90% and 94% of the time in influence-operation simulations, and that it tested whether models could autonomously plan multi-step influence campaigns. With safeguards enabled, the models refused nearly every task. Without those safeguards, Anthropic says only Mythos Preview and Opus 4.7 completed more than half of the tasks. That is a sober reminder that raw model capability and deployed model behavior are not the same thing.

Anthropic also published its evaluation methodology and open-source dataset, which may turn out to be the most important part of the post. Election integrity is a high-stakes domain where labs have often asked the public to trust internal testing they never show. By putting numbers, methods and benchmark materials on the record, Anthropic is nudging the discussion toward repeatable safety evidence. The company is also continuing product-side interventions, including election banners on Claude.ai that direct users seeking voting logistics to trusted sources such as TurboVote during the U.S. midterms.

There are limits here. A 95% or 100% score in an internal evaluation is not proof that real-world misuse disappears, and Anthropic says as much by promising continued monitoring and updates. But the direction is meaningful. As AI systems become part of how people search, debate and decide, election safeguards cannot stay at the level of brand messaging. They have to become measurable deployment practice. Anthropic’s post is one of the clearest examples this year of a frontier lab trying to show its work instead of asking for blind trust. The primary source is here.