Cloudflare lets you charge AI agents per request, settled in USDC
An HTTP 402 flow settles in USDC on Base, no checkout. Auto Router also lands to cut model spend ~30%, plus pay-per-crawl and Perplexity's new embedder.
Copy markdown
Pay-per-call for agents, no checkout page
Cloudflare's Monetization Gateway answers an agent with an HTTP 402 and a price; the agent signs and the request settles in USDC on Base through Coinbase's x402 Facilitator — no redirect, no separate payment API. You set prices per request by URL, header, or query, so you can ship a paid API or MCP tool that agents pay for on their own. Closed beta for US sellers and buyers.
Auto Router trims model spend up to 30%
Point AI Gateway at the model id `cloudflare/auto` and a multi-head classifier on Workers AI scores every request — 14 task types rated on complexity, ambiguity, stakes, and context — then routes to the cheapest model that still clears the quality bar, accounting for cache and model-switching costs on long sessions. Cloudflare measured up to 30% savings versus always calling a frontier model (benchmarked against GPT-6 Sol and Claude Opus 5.5). Free during the beta.
Pay Per Use turns your site into a toll booth
The same 402 rails power Pay Per Use, which lets publishers charge AI companies for content access instead of blocking their crawlers outright. If you run a content site, it's a metered gate rather than a flat robots.txt block.
Perplexity ships a context-aware embedder
pplx-embed-v2-context-9b is live on Hugging Face (preview; API coming) and tops ConTEB on nDCG@10. It retrieves the answer chunk plus its supporting context — 45.5% answer recall and 40.6% evidence recall at K=10 on Perplexity's context-bench — so an agent can verify a claim without re-reading the whole doc. Matryoshka dims from 256 to 2048, or int8, let you trade recall for cheap ~1KB vectors in a RAG stack.