The 'Half-Life' of New AI Models Has Shortened to One and a Half Months
Companies using paid AI services typically employ one model for about four weeks before switching to a new one. Vercel estimated the 'half-life' of current models at six weeks.
Vercel analyzed token transactions through its gateway and found that clients begin testing new AI models within days of their release. Code editors and specialized software frameworks simplify the transition, so switching models requires almost no redevelopment of the environment.
By September, approximately 17% of clients were working with models available for less than two weeks. Older solutions do not disappear entirely but are used less intensively, which also reduces token costs. According to Vercel, interest in a model noticeably wanes about six weeks after its launch.
In the paid segment, Anthropic's solutions are most frequently used. Companies often employ several models simultaneously, so the shares of individual products may overlap. One of the most notable models in terms of spending was Claude Opus 5.5: in the first full month of observations, it accounted for just over 14% of costs.
The use of tools for generating images and videos is also growing. Video generation models are now used by one in six companies involved in AI development, up from 10% previously. Among image generation tools, OpenAI's GPT-Image-2.5 Sunburst led as of October.
