Vercel: AI Models in Production Have a Six-Week Half-Life

By Nathan Brooks

AI models in production last about six weeks before half their users move on, according to Vercel's November 2026 AI Gateway Production Index. The deployment platform, which routes more than 100 trillion tokens a month between production applications and model providers, says enterprise adoption is far more fluid than lab benchmarks suggest.

The numbers are stark. More than half of October's language-model requests went to models less than 60 days old, up from under three in ten before September. A typical model peaks within a month of launch. Four weeks later, three in four teams that used it at peak have moved on—whether or not a successor shipped in the same family. Most stay on the gateway, just with different models underneath.

"AI models now have a half-life of about six weeks," Vercel writes. "That means that, to the teams using them, labs are only as good as their latest model and have to re-earn their business with every new one."

How fast is the churn, exactly?

The median model family ships a new version every 61 days—roughly every two months. Teams don't wait around: in September, about one in six language-model requests went to models less than two weeks old.

Among teams running at least 10 million tokens a month, roughly half changed their most-used model in October. Four in five of those switches went to a more recent release. Six in ten moved to an entirely different lab.

Older models don't disappear, but they shrink to a sliver. Seven in ten high-volume teams still ran a model at least ten months old in October. For the typical team, that model handled less than 3% of tokens—a fallback, not a workhorse.

What does this change for practitioners?

If your pipeline assumes a stable model, you're already behind. The real engineering challenge is no longer picking a model; it's building pipelines that swap models continuously without breaking downstream logic. Integration effort no longer buys lock-in, for vendors or for you.

Decision models take a quarter of requests

Production AI is splitting into two jobs: generation and judgment. Decision models—ones that classify, score, or pick between choices instead of generating text—cost roughly a seventieth as much per request as a language model.

They arrived on AI Gateway in mid-September. By October they accounted for 26% of all gateway requests while representing just 0.53% of spend. Nearly a third of paid teams used one, and about 11% used a decision model without sending a single request to a text-generating model.

Adoption scales with volume:

• Just over one in five teams under 10 million tokens use a decision model
• Nearly half at 100 million–1 billion tokens do
• Seven in ten teams above 10 billion tokens do

Almost all of that request volume went to Jev from TypeSafe AI, which alone received nearly four times Anthropic's total requests.

Voice adoption accelerates

About 11% of paid teams used a voice model in October, up from 4% in July—a 161% jump. Among teams that paid in both months, voice usage rose from 5% to 14%.

The share of teams sending at least 100 voice requests a month grew fivefold. Six in ten of September's voice teams returned in October, up from fewer than half in July. Speech-to-text remains the most common type, nearly tripling since July; text-to-speech has more than tripled. Teams are also shifting away from all-in-one voice models toward assembling separate speech-to-text, language, and text-to-speech components.

Anthropic wins teams; OpenAI wins wallet

Last December, 70 of every 100 paid teams used an OpenAI model and 48 used Anthropic. By October the two were nearly level at 55 and 54 per 100.

Anthropic's gain came largely from new teams: in October it was the first-choice lab for nearly twice as many new teams as OpenAI, and it has held that lead in nine of the last 11 months. Among the top 1% of teams by spend, more than nine in ten use both labs—so competition has shifted to share of wallet. OpenAI now takes nearly 15 cents of every dollar spent on the gateway, up from about 9 cents in December. Anthropic's share has halved to about 32 cents over the same period.

Image and video hit records

Image and video both posted their biggest month yet on AI Gateway. About a third of the top 1% of teams by spend now use an image model, up from a quarter in July. One in six use a video model, up from one in ten. OpenAI's GPT-Image-2.5 Sunburst led image spend in its first full month; ByteDance's Seedance 2.5 led video spend for a third consecutive month.

For labs, the takeaway is brutal: ship improvements faster than the half-life, or lose the next cycle's traffic entirely. For everyone else, treat every model choice as temporary by design.