The cost of processing Russian text varies significantly across AI models due to differences in tokenization algorithms. A study comparing 11 models, including GPT, Claude, Gemini, and others, reveals that the number of tokens required for the same Russian text can differ by a factor of five.
Claude Opus 5 requires 3.9 times more tokens to process Russian text compared to GPT-5.5, despite both models having the same price per token. This makes the effective cost of using Opus 5 for Russian-language tasks nearly four times higher than that of GPT-5.5.
Models optimized for Russian, such as YandexGPT and GigaChat, demonstrate the highest efficiency, with Russian text occupying fewer tokens than its English translation. In contrast, models like Claude Opus 5, DeepSeek, and Qwen show a higher token-to-character ratio for Russian, increasing costs for developers.
The research suggests that comparing models based on the price per million tokens is misleading for Russian-language products. Instead, developers should calculate costs based on the price per 1,000 characters or typical user requests to accurately estimate budgets.
Data for the analysis was collected on September 27, 2026, using median prices from aggregators. The author notes that service providers may have varying pricing, and usage statistics from API intermediaries should be verified against official developer data or local tokenizers.
