Volume 343, Issue 3 · 6 articles
Cloudflare launches a post-quantum Geo Key Manager in closed beta to keep decryption keys within chosen geographies.
Cloudflare is launching a post-quantum geographically restricted key manager in closed beta as part of its Data Localization Suite.
PQ GeoKey Manager wraps each decryption key so only Cloudflare servers in the allowed geography can decrypt it.
General availability is planned for Q1 2027.
Cloudflare says the launch supports data sovereignty needs of governments, financial institutions, healthcare and defense organizations while advancing its migration of all products to post-quantum cryptography.
Cloudflare and Commerce.com use WebMCP to give AI shopping agents structured tools on storefront pages.
Cloudflare says AI agents still struggle to shop online because they work from HTML, guessing at buttons in a slow, token-heavy and brittle process.
WebMCP, an experimental browser standard from Google, gives agents structured tools on a page, such as search_products, so they can call exactly what they need.
Cloudflare and Commerce.com describe how they are preparing storefronts to expose these tools.
The post explains how this makes agentic shopping faster, cheaper and more reliable.
Cloudflare cut the cost of its AI code reviewer by 68% to $0.77 per review while also speeding it up.
Cloudflare says the average cost of a review by its AI code reviewer rose from $1.19 in April to about $2.46 by late August without major architectural changes.
In mid-September it tasked agents with finding where tokens were going and cutting them aggressively.
Average review cost fell to $0.77, down 68%, and median review time dropped from 4m 09s to 1m 32s.
The post details the optimizations behind the savings.
Cloudflare's Clairic extracts data and estimates values in one pass with confidence scores, cutting a demo workflow's time 69%.
Cloudflare unveiled Clairic, a specialized System One model that extracts information from documents and images, performs basic calculations and estimates values without generating text.
In a customer service demo, combining Clairic with Cloudflare's Clef decision model and the GLM-5.3 LLM cut processing time 69% and costs about 49%, to $6,032 per million claims, though results are company-reported.
Clairic returns confidence scores and null values rather than unsupported answers, has a 65,536-token context window and costs from $0.10 per million input tokens with no output charges.
It is in preview on Workers AI.
Cloudflare previews Clairic, a fast System One extraction model hosted on Workers AI, with open weights planned.
Cloudflare is previewing a new class of System One-style models called extraction models, starting with Clairic, hosted on Workers AI as @cf/cloudflare/clairic.
Clairic extracts, computes, estimates and forecasts values from unstructured inputs, which Cloudflare says it does with higher accuracy, lower latency and lower cost than traditional LLMs.
The post references TypeSafe's Jev model and the idea of fast, precise System One models versus reasoning-focused System Two LLMs.
Cloudflare intends to release Clairic as open weights soon.
Cloudflare cut more than 100 TB of memory by using probability analysis to drop 90% of hash points in its Rust consistent-hashing library.
Cloudflare documented how it cut the memory footprint of its internal load-balancing service by more than 100 TB by optimizing pingora-ketama, the Rust consistent-hashing implementation behind its Pingora Backend Router.
Statistical analysis showed diminishing returns from more hash points, with the final 90,000 hashes improving distribution error by only 0.7%, so engineers cut hashes per backend by about 90% while preserving accuracy.
The birthday paradox explained rising collisions among 32-bit hash points.
They also packed each ring entry from 8 bytes to 6 and rolled out the new ring gradually to avoid invalidating cached content across the network.