The article reports four systems scoring between 67 and 72 of 76 checklist points across 14 Russian-prompt tasks. It also reports a pronounced failure by Nano Banana Pro on an 18-element table, while Nano Banana 2.1 lost capitalization and punctuation in poster and infographic tasks.
Mechanism: Aggregate binary checklists collapse heterogeneous capabilities—exact text, spatial arrangement, counting, and image editing—into one total, allowing task-specific weaknesses to disappear inside a close overall score. Why now: The comparison tests a newly released model against its predecessor, Pro, and GPT Image 2.5 on Russian-language tasks, exposing a language-and-task slice not covered by the cited external benchmark described in the article.
First implication: Select evaluation cases that mirror the intended production task rather than rely on a leaderboard total. Evidence boundary: One generation per task, one evaluator, no blind comparison, and no generation-time measurement; the reported results are not a general ranking. First watch condition: Whether repeated, independently scored runs preserve the same task-level error patterns. The article provides prompts and checklists for attempted replication, but this account does not independently validate the outputs.