Amid the latest technological advancements by global internet infrastructure giant Cloudflare, a transformation in "computing cost-effectiveness" is taking place. According to the Kimi Open Platform, Cloudflare has officially integrated Moonshot AI's open-source model Kimi K2.5 into its core production operations.

This decision was not made on a whim, but rather based on rigorous performance evaluations: Kimi K2.5 successfully replaced the previous high-cost closed-source models in programming and Agent tasks, thanks to its ultra-large 256k context window, keen visual input capabilities, and high stability in multi-round tool calls.

image.png

Cost reduction of 77%: A leap from millions to a fraction

Cloudflare engineers have widely used Kimi K2.5 as the "main driver" in daily development. A typical application case is an AI agent responsible for scanning code repository security flaws, which processes over 7 billion tokens per day. According to calculations, maintaining the original closed-source model solution would cost approximately $2.4 million annually for just this one use case.

After switching to Kimi K2.5, the inference cost dropped by 77%, reducing the previously heavy financial burden to just a fraction. More encouragingly, this rapid cost compression did not lead to any quality compromise; the model even accurately identified 15 confirmed security vulnerabilities in a single repository at once.

Embracing the open-source trend: Strategic transformation to reduce reliance on closed-source models

At SXSW2026, Cloudflare CEO Matthew Prince predicted that future internet traffic will be driven by hundreds of millions of AI agents, with inference demands growing exponentially.