Cloudflare AI Gateway

What people say

There isn't a dedicated review corpus for AI Gateway the way there is for consumer SaaS — it's developer infrastructure, so the signal comes from Cloudflare's own docs and changelog, comparison blog posts, and scattered developer commentary rather than G2 star ratings. That said, the shape of the feedback is consistent.

The pitch that lands: point your existing OpenAI, Anthropic or Workers AI calls at one gateway URL and get caching, rate limiting, retries with fallback, and a unified analytics dashboard — token counts, cost, latency, error rate — across every provider you use, with zero markup on inference (you pay the same per-token rate you'd pay the provider directly). It's free on every Cloudflare plan, which is the main reason teams try it: no procurement conversation needed to start.

Where it runs into limits is scale and sophistication. Reviewers comparing it to dedicated AI gateways such as Portkey, LiteLLM and Kong point out it lacks semantic caching — only exact-match requests get cache hits, so paraphrased prompts still hit the model. There's no MCP (Model Context Protocol) support yet. And every request adds a proxy hop, with published latency overhead in the 10-50ms range, which matters more for chat-style low-latency use cases than for batch jobs.

Cost surprises show up around logging, not inference: the free Workers plan caps at 100,000 AI Gateway logs a month across all gateways, and once you're pushing real production volume, you need a Workers Paid plan (raising the cap to 1,000,000) to keep full visibility — teams have reported this turning into a real line item at scale even though the gateway itself stays free.

Two 2026 additions worth knowing about: spend limits, which are hard dollar caps that block requests once a budget is hit (distinct from simple rate limiting), and Unified Billing, which lets you pay for third-party model usage through your Cloudflare invoice for a 5% convenience fee. Free DLP scanning for financial information and national ID patterns is also new and applies on all plans.

Bottom line from the available commentary: a solid default if you're already inside Cloudflare's network and want observability at no setup cost; teams running high-volume or heavily regulated AI workloads tend to graduate to a dedicated gateway product once semantic caching or MCP support becomes a requirement.

Summary of public user & expert reviews, compiled by RECATOOLS.