Guide
OpenClaw + GLM 5.3 FlashX: Flash at 5x Speed, 2.5x Price
On 18 September 2026 Zhipu released GLM 5.3 FlashX — not a new model, but GLM 5.3 Flash served on infrastructure tuned for throughput, at up to 200 tokens/s. Same weights, same benchmark profile, 2.5x the price. Here is what FlashX is, what the extra speed buys, and how to run it on an OpenClaw agent.
What GLM 5.3 FlashX Is
FlashX is the same mixture-of-experts model as Flash: 320 billion total parameters, about 18 billion active per token, natively multimodal — text, images and video in, text out — with a 1M-token context window and a 128K maximum output. It is the model that ran anonymously as ox-alpha before its August reveal, now with a faster serving tier.
The important thing to internalise: the name suggests a new generation, and it is not one. Zhipu published no new benchmark or architecture for FlashX, because the weights did not change. What changed is the infrastructure — more compute and inference optimisation behind the same endpoint. When you select FlashX, you are buying time-to-first-finished-answer.
- Up to 200 tokens/s output, which Zhipu describes as about 5x GLM 5.3 Flash. “Up to” is doing real work in that sentence; treat it as a ceiling, not a median.
- No capability change. Same 320B/18B base, same multimodal inputs, same 1M context and 128K output cap as Flash.
- 2.5x the price — $0.37 / $1.25 per million input / output tokens against Flash's $0.15 / $0.50.
- Not yet on the GLM Coding Plan. Z.ai's docs list Flash on the plan at three times the GLM 5.3 quota, and state FlashX is not yet available there.
Why There Is a FlashX at All
Flash landed harder than Zhipu expected. Since the ox-alpha test period it has been one of the busiest endpoints on OpenRouter, and Zhipu says the demand filled its inference cluster — it reports serving Flash traffic on a cluster of more than 100,000 domestically produced AI chips, with an Encode–Prefill–Decode disaggregated architecture that separates multimodal encoding, prompt prefill and token decoding into independently scaled pools.
Zhipu also says its own GLM 5.3-driven infrastructure agent helped engineers optimise kernels and the serving stack, and that the optimised stack improved end-to-end serving about 3x over its initial baseline on the same hardware. FlashX is the paid, latency-prioritised tier that came out of that work: instead of rationing the fast path, Zhipu prices it separately and charges for it. All of this is Zhipu's own account of its infrastructure rather than an independent audit.
The Pricing Trade
FlashX is priced at 2.5x Flash. Published Z.ai rates per million tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GLM 5.3 FlashX | $0.37 | $0.075 | $1.25 |
| GLM 5.3 Flash | $0.15 | $0.03 | $0.50 |
Cached-input storage on both is currently free for a limited time, and reasoning tokens bill as output like everywhere else on this family. The question is not whether FlashX is “worth it” in the abstract — it is the same intelligence at 2.5x the bill — but whether latency is the thing standing between you and a working product.
It usually is when a person is watching: a support agent, a live coding session, a tool loop that runs ten steps before it answers. At roughly 200 tokens/s a long answer arrives in a few seconds instead of ten or more, and that changes whether the thing feels usable. It is usually not when nobody is waiting: overnight research, batch document processing, scheduled jobs. There, Flash produces the same output for less.
Running GLM 5.3 FlashX Today
On Z.ai's own API the model code is glm-5.3-flashx. Two things decide how you reach it: FlashX is not on the GLM Coding Plan yet, and at publication OpenRouter listed the Flash base as z-ai/glm-5.3-flash but no FlashX slug. So there are two honest paths.
- Direct Z.AI API — the route that works today. Point a self-hosted agent at it with the config below, or set it wherever you can enter a model id. Use a standard pay-as-you-go Z.AI key, not a GLM Coding Plan key.
- Managed, on OpenClaw Launch — the dashboard model picker lists the models your connected providers expose, so GLM 5.3 FlashX becomes selectable here as soon as Z.ai's catalog or OpenRouter lists it. Until then, the same 320B/18B base model is selectable through an OpenRouter key as
z-ai/glm-5.3-flash, at 2.5x less. Bring-your-own-key selection needs a paid plan; the free trial runs on a fixed set of free models.
/model chat command. The model can change mid-conversation, and history is preserved.Self-Hosted Configuration
Running OpenClaw on your own server? Configure the direct Z.AI provider and point the default agent at FlashX in your openclaw.json:
{
"models": {
"providers": {
"zai": {
"baseUrl": "https://api.z.ai/api/paas/v4",
"apiKey": "your-zai-key",
"api": "openai-completions",
"models": [
{ "id": "glm-5.3-flashx", "name": "GLM 5.3 FlashX" }
]
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "zai/glm-5.3-flashx"
}
}
}
}Prefer OpenRouter? The same base model is there as openrouter/z-ai/glm-5.3-flash, and if a FlashX slug appears it will follow the same openrouter/z-ai/glm-5.3-flashx shape. Confirm the current ID on openrouter.ai/models before deploying rather than copying it from any guide, including this one — a stale slug surfaces as a model-not-found error.
Three Things That Will Surprise You
Thinking cannot be turned off. Reasoning is mandatory on the Flash family. The Z.ai API accepts only thinking.type: “enabled”, and OpenRouter reports reasoning as mandatory with a default effort of max. If you are used to disabling reasoning to cut latency on cheap models, that lever does not exist here — budget reasoning tokens as output, and note that they are exactly what FlashX is trying to deliver faster.
The recommended settings are the same unusual ones. Temperature 1, top_p 0.95, reasoning_effort: max, and thinking.clear_thinking: false. For streaming, enable both stream and tool_stream — and streaming is where a 200 tokens/s ceiling is actually visible, since a non-streaming call hides the win behind one long wait.
The Coding Plan does not include it yet. If you subscribe to the GLM Coding Plan rather than paying per token, Z.ai documents GLM 5.3 Flash at three times the GLM 5.3 quota but states FlashX is not yet available on the plan. On a Coding Plan, run Flash; use FlashX on pay-as-you-go until the plan catches up.
When to Pick Which
| Pick | When |
|---|---|
| GLM 5.3 FlashX | Interactive sessions, high-frequency agent loops, anything where a person or another agent is waiting on each turn and latency decides whether it feels usable |
| GLM 5.3 Flash | Everything else — the same intelligence for 2.5x less, including batch work, scheduled jobs and high-volume background agents |
| GLM 5.3 | You want the strongest GLM on hard coding and security review and the cost difference does not bind |
| DeepSeek V4 | Text-only workloads where you already have DeepSeek keys and pipelines |
Put simply: FlashX does not make your agent smarter, it makes it feel quicker. Buy it where that feeling is the product; stay on Flash where it is not.
What's Next?
- OpenClaw + GLM 5.3 Flash — the cheaper tier FlashX is built on
- Hermes + GLM 5.3 FlashX — the same model on Hermes instead of OpenClaw
- OpenClaw + GLM 5.3 — the flagship this family undercuts
- OpenClaw + GLM — the wider GLM-4 and GLM-5 lineup
- OpenRouter setup guide
- Compare all models