In September 2026, eight neural network models were tested for their ability to generate Russian text on images, including complex punctuation and specific formatting. The test involved creating memes and four-panel comics, with tasks ranging from simple captions to continuing existing visual narratives. Five of the eight tested models successfully processed all 56 words without errors.
Performance varied across models: Seedream 4.5 made one error, Qwen Image 2 Pro had three, and Flux 2 Max recorded nine errors. The cost per image generation ranged from 1.35 ₽ to 27.5 ₽, with processing times between 14 and 63 seconds. All tests were conducted via the BotHub aggregator, which allows access to these models through a single chat interface with payments in rubles.
The model gpt-image-2.5 emerged as the most cost-effective solution, producing a meme for 1.35 ₽ and a four-panel comic for 11.25 ₽, while maintaining perfect text accuracy. In contrast, Flux 2 Max was identified as the most expensive and least accurate model in the test group. The study provides detailed prompt templates and a checklist for users looking to generate text-heavy images.
