Cloudflare opens up account and network plumbing as Vercel floods AI Gateway with models

Cloudflare connects the network layer

The week's biggest infrastructure story is Cloudflare's general availability of Unified Routing for Cloudflare WAN and Magic Transit, now the recommended setup for all new accounts. It folds the standard Cloudflare One connectivity onramps together with Automatic Return Routing, BGP support, and custom client subnets — the kind of plumbing that used to require careful manual steering, now presented as the default path. Teams evaluating WAN or Magic Transit deployments should treat this as the baseline they configure against rather than an add-on.

Account management got more flexible too. Cloudflare now lets customers create additional Free accounts self-serve through the dashboard, spin up standalone Free accounts programmatically with user-owned API tokens or OAuth access tokens, and have Super Administrators mint up to five Free accounts inside an Enterprise Organization. For platform teams that treat accounts as disposable units of isolation, that removes a long-standing friction point.

Elsewhere in the platform, Cloudflare Rules added dry-run validation before ruleset deployment, catching bad expressions, action parameters, permission gaps, and quota limits without publishing anything — the dashboard does it automatically, and the Rulesets API takes a dry_run=true parameter. Logpush jobs can now exclude identified DDoS attack traffic across the http_requests, firewall_events, and network_analytics_logs datasets, which should make delivered logs far more usable during an attack. And in a heads-up for long-lived deployments, Cloudflare will stop publishing cloudflared builds for 32-bit Windows and Intel macOS in 2027, following Windows 10 and Intel Mac end-of-life. If you have tunnel fleets on that hardware, the clock is now visible.

Workers fills in observability and lifecycle gaps

Cloudflare's developer platform shipped a broad set of smaller improvements. Tracing got substantially better: Workers traces now follow JavaScript RPC calls across Worker and Durable Object boundaries, showing caller-side session and method spans that previously stopped at the RPC edge, with no code or SDK changes needed once tracing is enabled. Browser Run gained both typed session and DevTools methods on its bindings and an Inspect panel for session recordings with searchable console logs, a full network waterfall, and expandable DOM views exportable as JSON or HAR. Workflows can now delete instances individually or in batches of up to 100 via the API or Wrangler 4.125.0+, freeing stored state and stopping execution — welcome for anyone whose instance storage bill has been creeping up.

Smaller but useful: Workers AI added a rejectIfBusy option so synchronous inference fails fast instead of queueing when capacity is tight, Python Workers can now reach PostgreSQL and MySQL through Hyperdrive, R2 Data Catalog exposes per-table maintenance schedules and a manual compaction queue, and Super Administrators can invite teammates to a single Worker with one of four scoped access levels.

The Workers toolchain itself saw a steady stream of updates. Wrangler and the Vite plugin now build Dockerfile-backed Container images when experimental Build Output is enabled, and named Container images are now selectable explicitly through Wrangler config and the Vite plugin. Flagship bindings now evaluate flags against the local Miniflare store during development by default, keeping dev offline and isolated from production, with wrangler flagship flags pull to seed it. Wrangler also added placement configuration for Previews and fixed preview exports, while Miniflare aligned Local Explorer Workflow status with production and stopped intermittent assertion failures in synchronous D1 calls under load. On the runtime side, workerd enabled the Rust I/O backend by default and added an inbound UDP handler.

Vercel turns deployments into line items and Gateway into a model buffet

Vercel's most practical change is cost transparency: deployments now show billable duration and CPU minutes in the dashboard, vc inspect, and the REST API. Enterprise teams on Flexible Commitment plans can pair that with Spend Management budgets, alerts, webhooks, and optional pausing of production deploys. Build costs are now controllable per deployment, since Turbo build machines became opt-in on a per-deployment basis, and static artifacts deploy via the CLI in under a second.

AI Gateway had a busy week of model additions: Xiaomi's MiMo V2.6 Pro, Flash, and Pro UltraSpeed, Z.ai's GLM 5.3 FlashX, and Grok 4.7 at 40% off for a week, plus Jev through the TypeSafe-compatible and native HTTP APIs. Agents got new surfaces too: Microsoft Teams support in Vercel Connect, experimental WebMCP support in mcp-handler for in-browser agents with no separate OAuth flow, and skills installed straight from a Notion database. v0 can now install private npm packages using an NPM_TOKEN or NPM_RC variable, and Vercel Sandbox became a Harbor 0.22.0 eval environment for Terminal-Bench.

The AI SDK kept pace: mid-conversation tool discovery for code-mode landed in ai, the TypeSafe provider added experimental_evaluate for Choice, Score, and Boolean questions, Alibaba models now preserve reasoning across multi-turn requests, and the xai provider dropped the Chat Completions API entirely in favor of the Responses API — a breaking change worth checking before your next upgrade.