Cloudflare AI Gateway adds Auto Router and User Insights

Cloudflare’s AI Gateway beta Auto Router and free User Insights flag model overuse and route tasks to cheaper capable models.

On September 30, 2026, Cloudflare published two posts on AI Gateway: Auto Router in public beta and new User Insights views. Requests to cloudflare/auto are checked for format, credentials, policy, spend limits, and provider health, then a Workers AI model classifies them across 14 task types and four difficulty dimensions—complexity, ambiguity, stakes, and context. User Insights now covers tasks, model fit, cost, and conversation patterns, including an overkill view when a chosen model exceeds the task and a Potential Savings view where a faster or cheaper model could suffice. Auto Router applies those signals without a separate routing rule per workload; the capabilities are free for AI Gateway users, and reported categories include coding, research, writing, summarization, and data analysis. The posts agree that Auto Router is a public beta and do not conflict on the figures. Cloudflare says internal OpenCode use saved up to 30% versus frontier-only routing, and on a 291-trial internal knowledge-work benchmark cloudflare/auto succeeded on 86.6% of trials at $2.10 total, compared with Claude Opus 5.5 at 96.6% and $5.91 and GPT-6 Sol at 84.2% and $2.64.

• On 2026-09-30 Cloudflare launched Auto Router as a public beta in AI Gateway, reached via cloudflare/auto.
• Requests are filtered for format, credentials, policy, spend limits, and provider health, then a Workers AI model classifies them across 14 task types and four dimensions: complexity, ambiguity, stakes, and context.
• Updated User Insights, free for AI Gateway users, adds an overkill view and a Potential Savings view for model fit, cost, and conversation patterns.
• Reported task groupings include coding, research, writing, summarization, and data analysis; Auto Router applies model-fit signals without a separate routing rule per workload.
• Cloudflare says internal OpenCode use saved up to 30% versus frontier-only routing.
• On a 291-trial internal knowledge-work benchmark, cloudflare/auto scored 86.6% at $2.10 total, versus Claude Opus 5.5 at 96.6% and $5.91 and GPT-6 Sol at 84.2% and $2.64.