Artificial intelligence news.
Straight from the original announcements.

Vercel Flags SDK starts about 7x faster on cold starts, drops duration cost

October 29, 2026Source: Vercel
Vercel Flags SDK starts about 7x faster on cold starts, drops duration cost
Image: Vercel

Vercel has updated its Flags SDK so it initializes faster on cold starts, cutting p75 initialization time from over 70 ms to 10 ms in internal testing, according to the company. Usage reporting also no longer keeps functions running after the response is sent.

Key points
  • Vercel Flags SDK now loads the bundled flags datafile as soon as the client is created.
  • Each function invocation includes the current datafile revision, so no network request is needed to check it.
  • In Vercel's internal testing, p75 cold-start initialization dropped from over 70 ms to 10 ms, roughly 7x faster.
  • Usage events now go to the Vercel runtime, so reporting no longer extends function duration.
  • Update @flags-sdk/vercel to 1.5.3 or later and redeploy.

Two changes behind the faster cold starts

According to Vercel, two changes speed up SDK initialization when a function starts cold. First, the SDK now loads and parses the flags datafile that is bundled with a deployment as soon as the client is created. That means the data is ready before application code evaluates its first flag.

Second, every function invocation now includes the current datafile revision. With that information, the SDK can check whether the bundled datafile is up to date without making a network request.

Removing that network wait is what drives the improvement. In internal testing, Vercel reports that p75 cold-start SDK initialization time fell from over 70 ms to 10 ms, which the company describes as roughly 7x faster.

Flag reads after initialization

Vercel also gave figures for what happens once the SDK is running. After initialization, the company says, flag reads are effectively instant. Across cold and warm invocations, p99 flag-read latency is under 1 ms.

These numbers come from Vercel's own testing and have not been independently verified. Real results will depend on how an application uses flags and how often its functions start cold.

Usage reporting no longer adds to function duration

The second change concerns how the SDK reports usage. Previously it used waitUntil for this, which kept functions running after the response had already been sent. That extra time counted toward function duration.

The SDK now passes usage events to the Vercel runtime instead. As a result, usage reporting no longer adds to function duration, according to Vercel.

How to get the update

Developers need to update @flags-sdk/vercel to version 1.5.3 or later and then redeploy their application. Vercel gives this command for the update:

  • pnpm add @flags-sdk/vercel@latest

Because the improvements ship in the SDK, they take effect only after the package is updated and the app is redeployed. Teams that use Vercel Flags in functions that often start cold are the most likely to notice the change in initialization time. Further details are in the Flags SDK documentation.

Related news

Vercel Adds DeepSeek V4.1 Flash 1017 Ultrafast to AI Gateway
Products and toolsNovember 7, 2026

Vercel Adds DeepSeek V4.1 Flash 1017 Ultrafast to AI Gateway

Vercel has added DeepSeek V4.1 Flash 1017 Ultrafast to its AI Gateway through Fireworks, giving developers access to a model that can generate upwards of 1,000 tokens…

Vercel adds DeepSeek Search and Fetch to AI Gateway
For developersNovember 6, 2026

Vercel adds DeepSeek Search and Fetch to AI Gateway

Vercel has added DeepSeek Search and DeepSeek Fetch to AI Gateway, allowing developers to give models web search and page retrieval tools through one AI Gateway API key.…

Vercel Automatically Applies Token Pass Credit Across AI Gateway
Products and toolsNovember 5, 2026, 22:30

Vercel Automatically Applies Token Pass Credit Across AI Gateway

Vercel says Token Pass credit now applies automatically to included model usage through AI Gateway, regardless of whether a request uses an API key, an OIDC token, or fx…