AI NewsInfrastructureAnnouncement
Cloudflare launches Workers KV Instant in private beta, with p99 reads of 1.62 ms and a 60 percent price cut on reads
Cloudflare announced a new "instant" mode for Workers KV that reads a value in 1.62 ms at the 99th percentile and replicates writes to every edge in 256 ms, with reads priced 60 percent below classic KV.

Image: Cloudflare
Why it mattersWorkers apps that read feature flags or config on every request have been paying a cold-read tax for years, and this closes most of it: a request that used to wait 287 ms at p99 now waits under 2 ms, for less money.
A Workers app that reads a feature flag on every request has been paying a hidden tax for years: the first read in a region hits a cold cache and takes a quarter of a second. Cloudflare has put a new mode of Workers KV in private beta that cuts that read to under two milliseconds and makes every write reach every edge in a quarter of a second.
Cloudflare published the Workers KV Instant announcement on 1 October 2026. The new mode is powered by Quicksilver, a key-value store Cloudflare built in 2020 to carry its own configuration, and the company says nearly every request to Cloudflare already touches at least one Quicksilver key. KV Instant is the first time this store is offered to Workers developers as a product.
The numbers Cloudflare published
Cloudflare measured reads at 1.62 ms at the 99th percentile. The same read on classic KV at p99 is 287 ms, which the post calls a 100 times speed-up for the slowest reads. Writes reach every one of Cloudflare's more than 300 edge locations in 181 ms at the 95th percentile and 256 ms at the 99th. The numbers are Cloudflare's own, from the launch post, and the methodology behind them is not stated.
Reads cost $0.20 per million, down from $0.50 per million on classic KV, which Cloudflare labels a 60 percent cut. Writes and deletes cost $0.10 per operation, versus $5.00 per million on classic. Storage is $100 per megabyte per month, against $0.50 per gigabyte on classic. The classic prices and the Instant prices move in opposite directions for different reasons, which is the whole point of running two modes: Instant is paying to keep the hot path warm everywhere, classic is paying to keep a large data set cheap.
The limits that decide what fits
Each KV Instant namespace holds up to 10,000 key-value pairs, a key is at most 300 bytes, and the whole namespace tops out at one megabyte. Writes are capped at one per namespace per second. The two things that fit cleanly inside those limits are the two use cases Cloudflare names: feature flags and application configuration, where the whole table is small, every request reads it, and writes are rare.
A user session store does not fit. A product catalogue does not fit. Classic KV, R2 or Durable Objects remain the right place for those, and classic KV is not being deprecated. Cloudflare says the two modes are for different kinds of data, and existing namespaces stay as classic. To get the new behaviour, a developer creates a new namespace with "mode": "instant".
The private beta signup is at cloudflare.com/resource/workers-kv-instant-beta. No general-availability date is given, and the post names no migration path for an existing hot namespace that already lives in classic.
A Workers app that reads a flag on every request has, until today, had two realistic choices: pay the cold-read latency on classic KV, or build its own in-worker cache and accept the drift when the flag changes. Instant closes most of the gap on both sides at once: reads are fast enough that the in-worker cache is not worth the complication for a small table, and a change propagates in a quarter of a second rather than the better part of a minute. The size limits still rule out most of what Workers KV has grown to cover, so this is a new tool next to the old one, not a replacement.
Source
Primary source: Introducing Workers KV Instant, powered by Quicksilver, on the Cloudflare blog.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


