The performance of an AI model can decline significantly (by up to 33%) when information is provided across multiple messages. Several studies have found that accuracy drops even with newer models such as GPT-5 or Claude 4.6 when tasks are handled through a series of shorter prompts rather than a single prompt.
This loss in performance is not caused by memory or token limits. Instead, it appears to result from a kind of loss of attention: in longer conversations, models tend to draw conclusions too early, rely on their own earlier answers – which may already contain errors – or fail to give sufficient weight to information provided earlier in the conversation. Details mentioned in the middle of a dialogue are particularly likely to be lost. In addition, models tend to generate increasingly lengthy answers in long conversations, which may increase the risk of hallucinations. This decline in performance is referred to as context decay or long-chat degradation.
To counter this effect, researchers recommend starting new conversations regularly for complex tasks and summarising key information concisely. It can also help to combine important requirements in a single, structured and as complete as possible prompt rather than spreading them across multiple shorter messages.
By Ida Kandler
Source: The Decoder, 28 February 2026, by Matthias Bastian
