These LLMs are the best at resisting Russian propaganda

Open-weight models, including Nvidia’s Nemotron and Alibaba’s Quen, showed stronger results than Anthropic’s best models. GPT-5.4—OpenAI’s best-performing model—also performed relatively well on the benchmarks, providing “exemplary” responses on 54 percent of queries and achieving an 88.9 average score.

Not surprisingly, recent Borderlands models have shown a stronger tendency to resist Russian propaganda than models from a few years ago. Cloud 3.5 Haiku—the highest-rated model, released in 2024—only received an average rating of 73.1 on the benchmark. This mark would place it in the bottom third of models released in 2026 on this metric.

But this improvement over time was not uniform across all LLM manufacturers. Google’s most hype-resistant LLM, Gemini 2.5 Pro, is now almost a year old and has only reached an average score of 82 on the benchmark, largely due to its particular sensitivity to signals containing malicious words. The most recently tested Google model, Gemini 3.5 Flash, scored only 73 points on the benchmark, which is about the same as the Anthropic model released two years ago.

In a helpful post on the Propastop blog, the organization highlighted how many models showed little resistance to Russian propaganda when questioned in Russian. Google’s Gemini 3.5 flash received significantly lower benchmark scores in Russian than in English, as did open-source models like Moonshot’s Kimi K2 and Stepfun’s Step 3.5 flash.

What one country sees as propaganda, of course, another may see as a set of important cultural truths that LLM should support and reflect. A recent study by King’s College professor Gregory Asmolov analyzes how the Russian government – ​​through a recent technological alliance with other BRICS countries – is trying to influence AI models by introducing specific socio-political situations that are “culturally sensitive” to Russia’s viewpoint.