Cloudflare Enters Large Model Inference: Launches Kimi K2.5, Internal Security Agent Uses 7 Billion Tokens Daily with 77% Cost Reduction

Cloudflare recently announced that its Workers AI platform now supports large model inference. The first model launched is Kimi K2.5 from Moon Dark Side, featuring a 256K context window, multi-round tool calls, visual input, and structured output. The Agents SDK template has set it as the default model.

Cloudflare has internally applied Kimi K2.5 to daily development: engineers use it as the main programming agent in the OpenCode environment and integrate it into the automatic code review pipeline. One security audit agent processes over 7 billion tokens daily, finding more than 15 confirmed security issues in a single codebase. The cost has been reduced by 77% after switching to this model (the original mid-range commercial model cost about $2.4 million per year).

The platform also launched three improvements: prefix cache discount (no repeated charges for processed tokens in multi-round conversations), Session affinity header (route the same session to the same instance to improve cache hit rate), and asynchronous batch inference API (non-real-time requests can be queued and completed within 5 minutes).