Z.ai / Zhipu has split GLM-5.3-Flash into a faster paid tier. Official docs now list GLM-5.3-FlashX beside Flash: same multimodal Flash family, API model id glm-5.3-flashx, claimed peak decode around 200 tokens/s, and list prices roughly 2.5× Flash on the international pricing page. This is a latency SKU — not a second architecture drop.
Brocker already covered the base model when Flash shipped: 320B total / 18B active, hybrid sparse-linear attention, 1M context, native image/video/file input (GLM-5.3-Flash launch). FlashX docs reuse that stack. What changed on 18 Sep 2026 is product packaging: Zhipu says demand forced more infra and inference work on its Chinese-chip serving fleet, so it opened a higher-speed endpoint instead of shipping new weights or new benchmark tables.
International list prices on docs.z.ai (per 1M tokens): Flash $0.15 / $0.50 input/output versus FlashX $0.37 / $1.25. Cached input follows the same step-up ($0.03 → $0.075) — same ~2.5× step. Context stays 1M; max output 128K; multimodal inputs unchanged.
One operational caveat matters for Coding Plan buyers: Z.ai docs say Flash is fully available on the plan at 3× GLM-5.3 quota, while FlashX is not yet on the plan. Speed is pay-as-you-go API for now.
Confirmed
- Official Z.ai / bigmodel docs title the guide GLM-5.3-Flash/FlashX; model codes are glm-5.3-flash and glm-5.3-flashx.
- Vendor claim: FlashX inference up to ~200 tokens/s for faster responses; framed as infra + inference optimization on the Flash serving stack, not a new parameter count.
- docs.z.ai pricing lists FlashX at $0.37 input / $1.25 output per 1M tokens versus Flash $0.15 / $0.50 (cached input $0.075 vs $0.03).
- Coding Plan note in the same guide: Flash is live on the plan; FlashX is explicitly not yet available there.
- Shared Flash family facts restated in the FlashX docs: 320B-A18B MoE, hybrid sparse + linear attention, 1M context, multimodal coding / Office workflows — same capability surface as Flash.
Unknown
- Independent, sustained tokens/s under tool-heavy agent loads (not just peak decode) — vendor 200 tok/s is unverified here.
- Whether FlashX is bit-identical weights with different serving QoS, or a quantized / routed variant; docs do not publish a separate Hugging Face card for FlashX.
- When (or whether) FlashX joins GLM Coding Plan quotas, and how points would scale versus Flash.
- How often real sessions hit the marketed peak versus congested fleet averages after Ox Alpha / OpenRouter traffic spikes.
Our take
Do not file this next to a new open-weight drop. Flash already made the intelligence/price argument; FlashX sells time. At ~2.5× unit price for claimed ~5× decode headroom, the invoice math only works when latency is the bottleneck — interactive coding loops, streamed tool calls, live UI agents — not batch jobs that were already cheap on Flash. Same unit price ≠ same bill still applies the other way: a faster token meter that burns more speculative decoding can erase the feel-good speed if your harness retries. Cross-check the price ladder against the Brocker baseline at Chinese open models vs frontier API price/capability before treating FlashX as the new default.
