Teams that pay for AI models stay with a new one for an average of four weeks before moving to the next, according to Vercel's AI Gateway Production Index for November 2026. The company reads that as a half-life of about six weeks for any given model.
Vercel analyzed token transactions (the chunks of text that providers bill for) passing through its gateway. It found that teams adopt new models within days of launch. The report sorts adoption into 00-14, 15-30 and 31-60 day windows after release, and says early adoption has risen since June 2026. By September, almost 17% were using models that had been available for under two weeks.
Anthropic's Claude Opus 5.5, launched September 22, 2026, was the top model by spend in its first full month, October, with just over 14%. Among the most widely used paid models, Anthropic dominates: Sonnet 5.5 at 26%, Haiku 5.5 at 22%, Haiku 4.5 at 20% and Opus 5.5 at 18%. Those figures overlap because many teams run several models. Elsewhere, one in six AI development teams now use a video model, up from 10%, and OpenAI's GPT-Image-2.5 Sunburst led image generation spend in October.
The pattern follows from how cheap switching is. In a code editor or tool wrapper, pointing at a new model is a small change, so little holds a team to last month's choice. That makes release cadence a competitive weapon for labs: a model gets a few weeks of concentrated attention, and a rival's next launch resets the clock. For teams, constant switching means constant re-testing, a cost the data does not measure.
"Half-life" also deserves a light touch. Vercel says older models are retained to some degree, just with less usage and spend, and Haiku 4.5 at 20% fits that. The term describes fading attention rather than retirement. The numbers also come from traffic through Vercel's own gateway, so they describe the teams that use it, and it is an open question how closely the wider market matches.