Cloudflare brings Z.ai's GLM-5.3 to Workers AI

Cloudflare has added Z.ai's GLM-5.3 to Workers AI, making the model available through a globally distributed inference platform and an OpenAI-compatible endpoint. The August 28 release gives developers another hosted route to a 1,048,576-token reasoning model built for long-running coding and tool-driven agent workflows.

The change is about access and deployment, not a new model announcement. GLM-5.3 launched on August 18, but Cloudflare's integration lowers the operational barrier for teams already using Workers, AI Gateway, or Cloudflare's serverless stack.

What Cloudflare adds for GLM-5.3

Workers AI exposes GLM-5.3 through the env.AI.run() binding, REST API, OpenAI-compatible API, and AI Gateway. Cloudflare lists function calling and reasoning support, with reasoning enabled by default. The model accepts up to roughly one million tokens of context, which is useful for repository-scale analysis, long tool traces, and applications that need to retain large working histories.

Cloudflare prices the model at $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens. The model is not included in standard Workers Free billing: users need a Workers Paid plan or prepaid AI Gateway credits.

GLM-5.3's coding claims need context

Cloudflare reproduces Z.ai's reported benchmark results, including 88.2 on Terminal-Bench 2.1, 28.3 on Terminal-Bench 3.0, and 42.5 on SWE-Marathon. Those figures describe the model's intended strengths, but they are vendor-reported comparisons and should not be read as a guarantee for a particular repository or agent loop.

Artificial Analysis independently evaluates GLM-5.3 at about 60 on its Intelligence Index and records a one-million-token context window, roughly $1.40/$4.40 list pricing, and high output volume during evaluation. That external view supports the model's positioning while adding a practical warning: always-on reasoning and verbose outputs can raise per-task costs even when the per-token price looks competitive.

Deployment and safety trade-offs

The model's open-weight distribution is documented by Z.ai's official Hugging Face repository, which contains the checkpoint files and deployment instructions for Transformers, vLLM, SGLang, and Docker Model Runner. Hosted Workers AI access is operationally simpler than self-hosting a checkpoint listed at roughly 756 GB, but it does not remove the need to review data handling, tool permissions, and spend controls.

GLM-5.3 also carries public-interest security implications. Cloudflare highlights Z.ai's reported CyberGym result of 84.5, reflecting strong vulnerability-discovery capability. That can help authorized defenders, but teams should keep the model inside approved environments, limit tool access, log actions, and avoid treating benchmark scores as a safety assessment.

The practical availability decision

Workers AI is now a viable hosted option for developers who want GLM-5.3 inside Cloudflare's application platform and can accept paid access. OpenRouter lists multiple providers, while direct or self-hosted deployment remains possible through the official checkpoint. Teams comparing routes should measure complete task cost, latency, data residency, tool-call reliability, and the safeguards around autonomous execution—not just context length or benchmark rank.

For agent-building workflow ideas, see LinkLoot's AI agent tools guide. The next decision is operational: whether Cloudflare's integration fits the team's existing Workers architecture and governance controls well enough to justify moving long-running coding tasks onto a model with always-on reasoning.