Cloudflare AI Gateway adds a Web Search API for your agents
Birthday Week also ships a self-serve OHTTP gateway, AI Search GA with visual + OCR, 1.6ms Workers KV reads, and Kafka-style streams on R2.
Copy markdown
Web search, built into the inference call
AI Gateway now routes to Exa, Linkup, or Ceramic.ai so your model grounds answers in live results instead of its training cutoff. Call it over REST or a one-line Workers binding (env.AI.websearch()) at partner pricing with no markup, BYOK supported, and every query lands in your Gateway logs.
Spin up your own OHTTP gateway
A closed beta lets you enable a self-serve OHTTP Gateway as a paid add-on to a zone, so your backend receives requests without ever seeing the user's IP. The old Privacy Gateway is renamed OHTTP Relay to keep the two straight.
AI Search hits GA and learns to see
Cloudflare's managed retrieval pipeline (Workers AI + Vectorize + R2) is now GA and embeds raw image pixels via Qwen3-VL for real visual search, runs OCR on scanned PDFs, and takes files up to 10 MiB (was 4). Point it at a bucket and query across text and images.
Workers KV Instant kills the cold read
A new KV mode on Quicksilver returns a 1.62ms p99 read (vs ~287ms classic) and replicates 99% of writes in ~250ms, with no cold-read penalty on the first hit. Reads cost 60% less; writes and storage cost more, so it's aimed at config and flags you read constantly and change rarely.
K2 brings ordered event streams to R2
K2 is a serverless, durable, partitioned log built on R2 object storage, giving you Kafka-style ordered streaming with compute and storage split apart and no broker clusters to run. Good for event pipelines and agent task queues you don't want to operate.