Mutual self-assessments of general characteristics of some popular free AI tools

Abstract

Currently, there is a growing number of publications devoted to the analysis, evaluation, and optimal selection of AI models for specific applications, as well as market offers for services comparing and selecting AI tools for specific tasks. Along with this mainstream information, it is interesting to review the capabilities and characteristics of tools based on AI models’ «opinions» about themselves and their «colleagues». The purpose of this article is to conduct an experimental comparative review of model characteristic assessments and identify possible trends in them. The current generation of popular, accessible intelligent assistants, aimed at a wide range of users and their tasks, have been chosen for review. In particular, mutual assessments of the characteristics of ChatGPT versions 3.5+DALL-E, 4o, and 5, DeepSeek V3, Gemini 1.5 and 2.5, Claude Sonnet 3 and 4 are obtained and discussed. Cross-referenced 5-point self-assessments of current model generations according to specified criteria and their assessments of the expected capabilities of ChatGPT-5 are provided and analyzed. The advantages of models are discussed in the context of balanced assessments and recommendations for their application. Ambiguity, inconsistency, and possible deviations from objectivity have been identified. Models often demonstrate bias by overestimating their own capabilities. Among the possible reasons for this (obsolescence of some knowledge bases, some kind of hallucinogenicity, etc.), the most likely is considered to be the competitive and marketing orientation of models from different manufacturers. Therefore, when using the models themselves for relevant assessments and selection, it is essential to weigh their opinions collectively and exclude self-assessments. The emergence of GPT-5 is noted for its significance in changes to the «competitive landscape» and approaches to model evaluation criteria.