AI: what does Vercel’s six-week ‘half-life’ really measure?

Vercel observes a high turnover in the models used by teams using AI Gateway. Its figures describe adoption on one platform, not technical obsolescence after six weeks.

On AI Gateway, teams quickly reallocate their requests between AI models. The November 2026 edition of Vercel’s Production Index describes this turnover based on traffic on its platform in October. Its ‘half-life’ formula of six weeks refers to observed usage, not to models that stop working on a fixed date.

The term ‘half-life’, used by Vercel, gives a striking label to a measure of adoption. It does not mean that a model stops working after six weeks. To understand what the figure reveals, we need to look at who is counted, how versions are distinguished, and what teams continue to use after switching their primary model.

A snapshot of an intermediary’s traffic

AI Gateway acts as an intermediary between applications and model providers. Vercel says that more than 100 000 trillion tokens—units used to count the content processed by models—pass through it each month. Its report aggregates anonymised platform traffic up to October 2026. It therefore offers a view of the teams using this tool, not a survey of the entire AI market.

The report measures the age of a version from the first day it appeared on AI Gateway. A version is therefore counted separately from another version in the same family. This clock describes its presence on the platform, not necessarily the date on which its developer publicly launched the model.

In October, more than half of the requests to language models on AI Gateway concerned versions that had been available for less than 60 days. Before September, this proportion had never exceeded three in ten, according to Vercel. The figure indicates that recent versions account for a large share of the observed traffic. On its own, it does not show that they perform the requested tasks better.

What the six-week ‘half-life’ measures

Another statistic tracks teams after a version’s usage peak. In the month following that peak, three in four teams that had been using it stop doing so, Vercel reports. They may have moved to a newer version, another developer, or another available option. The report notes that most of these teams remain on AI Gateway: they therefore do not disappear from the sample when they change models.

The report puts the usage peak for a typical version in the month after its arrival. The turnover measured afterwards provides information about the renewal of usage; it does not establish that the version has suffered a technical loss of performance.

The figure for changes to the primary model relates to a different group. Vercel compared teams that processed at least 10 million tokens in September and October, and defines the primary model as the one receiving the most tokens. In this subset, around half changed their primary model between the two months. Of the teams that changed, four in five moved to a newer version and six in ten to another developer. These proportions therefore do not describe all teams, nor exactly the same phenomenon as stopping use of a version after its peak.

Changing your favourite does not mean abandoning everything

Older versions continue to be used alongside newer ones. Among teams processing at least 10 million tokens per month, seven in ten still used at least one model in October that had appeared on AI Gateway ten or more months earlier. For a typical team in this group, however, this older model accounted for less than 3 % of the tokens processed.

The distinction matters: a version can remain technically accessible and feature in a team’s activity without retaining a significant share of its traffic. The primary choice can change while some secondary uses remain in place. For an organisation, this coexistence may reflect different applications or functions that have not all been moved at the same pace; the published data do not make it possible to attribute a specific motivation to each change.

The report also shows that adoption is not limited to text generation. Decision models classify, assess or choose between options rather than produce a written answer. They accounted for 26 % of AI Gateway requests in October, but just 0,53 % of estimated spending. Nearly a third of the platform’s paying teams used at least one during the month. The share of requests therefore does not mean that a third of tasks or spending is attributable to them: teams may send many because these responses are short and relate to specific operations.

For teams, novelty creates a selection workload

This turnover makes the work surrounding adoption more visible: determining which version suits each task, checking the result in the product concerned, then deciding whether the gain justifies a change. Vercel’s data describe movements in traffic, not their causes. They do not say whether a team switched to improve quality, reduce latency, control costs or meet another constraint.

Vercel estimates spending using published list prices, which may differ from actual bills. The report therefore does not provide a direct measure of the cost actually paid by each organisation. And since the gateway operator observes the models available on its own service, its figures are useful for understanding its customers’ usage, but do not constitute a universal ranking of models.

Finally, this voluntary switching must be distinguished from withdrawals decided by providers. OpenAI, for example, defines deprecation as the process leading to the withdrawal of a model or endpoint; once the shutdown date is reached, the service concerned is no longer accessible. The documentation provides for customer notification and a withdrawal schedule. This process is distinct from a team’s choice to route its requests elsewhere.

For a professional team, the question is to check what a migration changes in its application: response quality, waiting time and actual cost. The movement in traffic observed by Vercel invites teams to examine these choices; it does not remove the need to assess the results task by task.