Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- The Deno team is joining Cloudflare and will end Deno runtime development after a year
Deno announced that the entire Deno team is joining Cloudflare to work on a shared platform built on the Workers programming model and Durable Objects, which Ryan Dahl says suit agent harnesses. Deno will support the Deno runtime for another year with monthly bug-fix and security releases, after which it will end its development of the runtime; Deno stays open source. Deno Deploy will keep operating for six months before shutting down, with migration help for paying customers moving to Cloudflare Workers, and JSR will keep running on Cloudflare infrastructure.
- Cloudflare's Clef decision models arrive on Ollama
Cloudflare's Clef decision models are now available on Ollama, per the Ollama account — a one-line "ollama pull clef" brings the local decision-model stack to any machine.
- Cloudflare open-sources Clef decision models, API-compatible with Jev and tops its index
Cloudflare open-sourced Clef and Clef-flash, classifier-style decision models that output typed answers with probabilities instead of free-form text, under Apache 2.0 on Hugging Face and hosted on Workers AI. Clef builds on a frozen Qwen3.8-27B backbone and Clef-flash on Qwen3.5-9B, using a prefill-only inference pass with parallel schema scoring rather than autoregressive generation, with vision input and a 64k context window versus Jev's 32k. The API is compatible with TypeSafe's Jev, making swaps drop-in, and Clef tops the Jev Decision Index: median latency is 209.3ms for Clef and 38.8ms for Clef-flash against 524.1ms for Jev, and Clef beat Jev in three of four Typesafe workflow evals. Training used rank-256 LoRA adapters, label-smoothed cross-entropy with Brier loss, and a custom RL method called RLCD. Cloudflare also announced an RL fine-tuning service built on AI Gateway, Workers AI and Containers sandboxes.
- Cloudflare rebuilds Containers for on-demand agent sandboxes with snapshots and 6x faster starts
Cloudflare shipped a public beta of a new durable_object scheduling policy for Containers, aimed at agent workspaces created on demand: code now picks each sandbox's image and compute size at start time, and filesystem snapshots via snapshotContainer() let sessions resume or fork from one checkpoint. ComputeSDK's independent benchmark put median time-to-interactive at 648ms, down from 4.049 seconds, and a burst test started 100,000 containers in about 5.4 seconds. A new cloudflare/debian-trixie image (Debian Trixie Slim plus Node.js 24 LTS) lets agents start without building their own. Cloudflare says rollouts are now just code, with canary, pin, migrate and rollback per sandbox. The legacy Container and Sandbox classes stay supported through December 31, 2026.
- Cloudflare's AI Gateway Auto Router picks the model per request, free in public beta
Cloudflare's AI Gateway launched Auto Router in public beta: set the model to cloudflare/auto and a multi-head classifier on Workers AI GPUs scores each request across 14 task categories plus complexity, ambiguity, stakes and context dependence, then routes to the model with the best expected quality minus an adaptive cost penalty. It is free during the beta. In Cloudflare's internal benchmark of 97 tasks, cloudflare/auto succeeded on 86.6% of them at $2.10 total, versus 96.6% at $5.91 for Claude Opus 5.5 and 84.2% at $2.64 for GPT-6 Sol — the company notes these are its own results, not independent conclusions.
- Cloudflare opens agent monetization betas: x402 Monetization Gateway and Pay Per Use
Cloudflare put two agent-commerce products into beta. Monetization Gateway (closed beta, US sellers first) lets domain owners charge AI agents per request, query or token for APIs, MCP tools, data feeds and websites: an unauthenticated request gets an HTTP 402 with payment instructions inline, authorized and settled in USDC on Base through Coinbase's x402 facilitator. Pay Per Use (beta) pays publishers when AI products actually use their content — buyers define what counts as a paid use and self-report each use via a JSON API, and Cloudflare verifies, bills and pays out monthly. Pay Per Use charges for downstream use, unlike 2025's crawl-based Pay Per Crawl, and publishers can combine both models.