Cloudflare Workers AI now runs Kimi K2.5
Cloudflare's large-model Workers AI post is useful because it connects model choice, inference cost, and agent infrastructure in one practical story.
A security agent processing 7 billion tokens a day on a proprietary model costs $2.4 million a year. The same workload on an open model through Workers AI costs 77% less. At agent-scale volumes, model choice becomes a budget decision.
Source
Powering the agents: Workers AI now runs large models, starting with Kimi K2.5
By: Cloudflare
Published: Mar 19 2026
Type: Release announcement
This isn’t just a model launch post. It connects inference cost, agent infrastructure, and the practical reality that when every employee has multiple agents running all day, proprietary model pricing stops working. That framing is more useful than another benchmark comparison.
Why agent infrastructure matters more than model benchmarks
Workers AI has served models since launch, but historically focused on smaller ones. Open-source LLMs lagged far behind frontier labs for a long time. Models like Kimi K2.5 changed that equation — frontier-scale reasoning at open-source pricing — but serving a model this large required real changes to Cloudflare’s inference stack.
Kimi K2.5 lands on Workers AI with a 256K context window, multi-turn tool calling, vision inputs, and structured outputs. Those are the features agent loops actually consume. The model sits inside the full Cloudflare developer platform: Durable Objects for state persistence, Workflows for long-running tasks, Dynamic Workers and Sandbox containers for secure execution, and the Agents SDK for orchestration. The whole agent lifecycle runs on a single platform instead of stitching together separate model API, compute, and state services.
The cost story is the grounding detail. Cloudflare’s internal security agent processes over 7 billion tokens per day using Kimi K2.5 and has caught more than 15 confirmed issues in a single codebase. Running that same agent on a mid-tier proprietary model would cost an estimated $2.4 million per year. Workers AI cut that by 77%. When personal agents and coding assistants run continuously for every engineer — and Cloudflare’s engineers are already there — the volume of inference makes cost the primary blocker to scaling, not model quality.
Behind the scenes, serving a model this large requires custom kernels on Cloudflare’s proprietary Infire inference engine, data/tensor/expert parallelisation, and disaggregated prefill (separating prefill and generation stages onto different machines for better throughput). These optimisations take dedicated ML engineering work that most teams don’t have. The platform argument is that you call an API instead of becoming a serving engineer.
Three platform features shipped alongside the model. Prefix caching avoids redundant prefill work in multi-turn conversations. Each new turn typically adds a few lines to the same system prompt, tools, and context. Caching input tensors from the previous request means only the delta goes through prefill, cutting time-to-first-token and improving throughput. A new x-session-affinity header routes consecutive requests to the same model instance for higher cache hit rates, and cached tokens now get a price discount.
Asynchronous APIs handle volume that exceeds synchronous rate limits. Instead of failing with out-of-capacity errors, requests queue durably and execute when headroom appears — a pull-based system that monitors GPU utilisation in real time and gives synchronous requests priority while still processing the async queue. Internal testing shows async requests typically execute within 5 minutes. This fits non-real-time workloads like code scanning agents and research agents.
The combination matters. A single agent call is cheap anywhere. Thousands of agents calling models continuously, with multi-turn context, tool definitions, and full codebases in the prompt — that’s where the stack around the model decides whether the system is practical. Prefix caching reduces repeated token processing. Session affinity keeps cache hit rates high. Async APIs absorb traffic spikes. The model is one layer; serving infrastructure is another.
My take
The shift this represents is subtle. Two years ago, the question was “which model is smartest?” Now it’s “which platform makes the model affordable and reliable at the volume my agents produce?” Cloudflare’s answer is vertically integrated: the model, the serving optimisations, the execution primitives, and the routing all live on the same network. That integration is a genuine advantage (no cross-cloud hops, same-network latency for Durable Objects and inference), but it’s also a lock-in bet. Teams should benchmark not just model quality but the full cost curve: tokens processed, cache hit rate at their actual conversation length, async queue latency under real load, and what happens when they need a model Workers AI doesn’t host yet.
Where this is useful
- When building agents that run continuously (coding assistants, security scanners, research agents) — the cost difference between proprietary and open models at sustained volume is the deciding factor, and this shows how to get frontier reasoning at a fraction of the price.
- When multi-turn agent conversations repeat large context (system prompts, tool definitions, codebases) — prefix caching with session affinity directly reduces cost and latency.
- When agent workloads produce traffic spikes that exceed synchronous limits — the async API turns capacity errors into queued requests that execute durably.
Apply this: Calculate your agent’s daily token volume and compare the cost on a proprietary model vs. an open model on a managed platform. If the gap is large, the model switch is a higher-leverage optimisation than prompt engineering.
What I want to remember
Remember: If I build agents on a cloud platform, compare the whole stack — model quality, serving cost, prefix caching, async capacity, and execution primitives — not just benchmarks. A cheaper model with good routing, logs, and stable execution can be more useful than a stronger model that’s painful to operate.