Z.ai's GLM 5.3 FlashX arrives on Vercel AI Gateway at ~200 tokens per second

Vercel has added GLM 5.3 FlashX to AI Gateway, giving developers a high-speed serving option for Z.ai's multimodal coding model behind a single API surface. According to a Vercel changelog post, the variant delivers inference at roughly 200 tokens per second, which translates into faster streamed responses for anyone waiting on generated output.

What FlashX is tuned for

Vercel frames the speed gain as most relevant to coding agents, tool loops and interactive applications — settings where a person, or an automated pipeline, is blocked until the model finishes responding. Coding agents are a natural fit because they tend to make many sequential model calls: draft code, invoke a tool, read the result, draft again. When each call streams back faster, the whole loop tightens, and the perceived responsiveness of the agent improves.

The underlying model is multimodal, per the announcement, although the post does not detail which input modalities the FlashX tier supports beyond its coding focus.

How to switch it on

The model is exposed under the identifier zai/glm-5.3-flashx and works across AI Gateway's supported API formats. For agent use, Vercel points developers to its coding agents guide and recommends running vercel ai-gateway setup, which creates a key and configures supported agents; you then pick zai/glm-5.3-flashx from inside the agent itself. Developers who want to evaluate the model before wiring it into a project can try it in the model playground or browse the full catalogue of language models available through the gateway.

What the gateway layer adds

AI Gateway is positioned as more than a simple proxy. According to Vercel, it provides one API for calling models, tracking usage and spend, and configuring retries, failover and performance optimisations that the company says push uptime above what the underlying provider offers on its own. It also ships with built-in custom reporting, budgets scoped to individual API keys, and routing rules for steering traffic between models.

On cost, Vercel says the gateway passes provider pricing through without adding a markup and takes no platform fee on inference — including on Bring Your Own Key (BYOK) requests, where teams use their own credentials for the underlying provider.

Why it matters

Latency compounds in agentic workloads. A coding agent that runs dozens of model calls per task feels every extra second of per-call overhead, so a serving tier tuned to around 200 tokens per second attacks a real constraint rather than a cosmetic one. Because FlashX sits behind AI Gateway, teams get it alongside failover, usage tracking and per-key budgets instead of building a separate integration, and the no-markup pricing means there is no cost penalty for routing through Vercel. For developers weighing alternatives to the incumbent coding models from the largest labs, an easily reachable, fast-serving option with a one-command setup lowers the barrier to running a genuine comparison — and for anyone already building agents on Vercel, it is close to free to try.