Generative artificial intelligence systems are increasingly appearing in decision-support processes, so it is essential to address the extent to which these models are prone to human-like cognitive biases. This study investigates whether anchoring bias can be detected in large language models in a simulated market decision-making situation where the AI's task was to determine the launch price of a new smartwatch. In the experimental setup, five different generative AI models (GPT-4o, GPT-4.5, Gemini 2.5 Pro, Grok 3 Beta, Sonar) were tested with low and high numerical anchor values. For each model, 20 runs were performed under both conditions (a total of 200 queries), and the results were analyzed using an independent sample t-test and an anchoring index. Based on the results, the GPT-4o model showed significant anchoring bias (AnI=13.12%), while in the case of GPT-4.5, this was more moderate (AnI =5.67%). The responses of the other models were completely consistent, with a standard deviation of 0 and no changes observed between different anchoring conditions. The hypothesis tests confirmed that the anchoring effect is not universally characteristic, but rather a model-dependent behavioral peculiarity. The study contributes to the measurement of bias sensitivity in artificial intelligence and to the development of a possible future behavioral benchmark. The practical significance of the research is that it draws attention to the fact that individual models may be sensitive to irrelevant numerical contexts, which can lead to biased results in business decisions. Therefore, for companies, not only language performance but also this type of bias profile may be an important consideration when selecting and deploying generative AI systems.