Cloudflare lifts Worker size limits while Python frameworks land

Bigger Workers, bigger ambitions

The week's defining story is Cloudflare's removal of Worker bundle size limits in favor of a single 64 MiB uncompressed cap across all plans. Previously, Wrangler's compressed-bundle check rejected deploys over 3 MB on the Free tier and 10 MB on Paid. That friction is gone: heavier frameworks, larger dependencies, and more substantial codebases now fit without hitting an arbitrary ceiling. The shift from compressed to uncompressed measurement is also notable — it aligns the platform with what developers actually reason about when bundling, and the --dry-run --outdir flag gives a straightforward way to audit bundle size before a deploy.

Complementing that expansion, Cloudflare Python Workers now support WSGI and ASGI frameworks. Django, Flask, FastAPI, and Starlette apps can run on the platform using the new wsgi and asgi exports from the workers module, with entrypoint() handling the default wiring and WorkerEntrypoint available for finer control. This is a meaningful unlock for Python developers who previously had to adapt their framework or hand-roll request handling — now a familiar application structure can be lifted onto the platform directly.

One cautionary item arrived the same week: Cloudflare D1 began enforcing free-tier daily query limits on September 1. Accounts exceeding daily row read or write limits now see query failures until the midnight UTC reset, with email alerts on limit reach. Stored data remains unaffected, so the blast radius is operational, not destructive. The fix guidance — add indexes, or move to a paid plan — is worth acting on before hitting a wall mid-project.

AI Gateway keeps adding models

Vercel's AI Gateway had a busy stretch on the model front. Meta's Muse Spark 1.3, Google's Gemini 3.8 Flash, Alibaba's Qwen 3.8 Max 0902, and Anthropic's Claude Fable 5.1 all joined the gateway, each accessible through a single API key with fallbacks, spend tracking, and traces on every request. There's also a limited-time deal: GLM-5.3 served by DigitalOcean at 50% off through September 7 via the promo model name. The pattern is consistent — AI Gateway continues to position itself as the consolidation layer for model access, removing per-provider accounts from the critical path.

Under the hood, the AI SDK v7.0.89 expanded the FilesV4 interface with optional metadata, streaming download, and delete operations, plus streaming uploads and abort signal support. The OpenAI and xAI providers picked up those same file operations, and the OpenAI provider added GPT-6 reasoning configuration updates. This rounds out the files story across the SDK ecosystem, giving providers a common interface rather than bespoke implementations.

Platform hardening and developer experience

Several releases focused on making day-to-day development less noisy. Wrangler 4.128.0 moved binding utilities into @cloudflare/workers-utils, shared the code with the Vite plugin and other consumers, added --date-start/--date-end filters for workflow instance listing, and trimmed its installed size by roughly 9.2 MiB. Query string redaction for observability arrived in Workers configuration, stripping query strings from request URLs in logs and traces — a useful privacy default for production debugging. The Workers Vite plugin now surfaces Worker export type fetch errors with the Worker name, HTTP status, and response body instead of a generic JSON parse failure, which should cut down on fruitless debugging sessions.

On the tooling side, Vercel added Basic build machines for Pro and Enterprise — a 2 vCPU/8 GB option for smaller apps, with Elastic still the default for new projects. Downloads of system updates aren't strictly part of this week's story, but the staggered canary Next.js releases continued with a fix for route info segment overrides in the dev overlay and adapter route collapses enabled by default.

The AI Gateway model additions and the SDK file-operation overhauls point to the same conclusion: this was a week where platform boundaries got wider — bigger bundles, more frameworks, more model access — and the free-tier enforcement at D1 is the reminder that with broader capability comes clearer limits.