A study by a user named donseo on the Habr platform reveals that processing Russian-language text in Claude Opus 5 is nearly four times more expensive than in GPT-5.5. The discrepancy arises from the models' tokenizers, which break down text into fragments differently before processing.
The research indicates that Claude Opus 5 identifies 390% more tokens for the same Russian text compared to GPT-5.5. Because billing is based on token count, users of the Anthropic model face significantly higher costs when working with Russian content, despite identical pricing per token.
The analysis covered 11 neural networks, including Gemini, DeepSeek, Qwen, GLM, Kimi, Grok, and Russian models like GigaChat and YandexGPT. While English text is tokenized similarly across most models, Russian text shows significant variance, with Claude Opus 5 often treating individual Russian characters as separate tokens.
In contrast, Russian-developed models demonstrate higher efficiency with Russian text, where it consumes fewer tokens than English. The study also noted that the imbalance for Russian text in models like GPT-5.x/6, Gemini, Grok, and GLM ranges between 18% and 23% compared to English.
Beyond technical costs, Anthropic continues to restrict access to Claude in Russia. The company has been actively banning accounts identified as using VPNs or other circumvention tools, with mass account closures reported throughout 2026.
