Zhipu rolled out GLM-5.3-FlashX on September 18, with API access opening at the same time under the model ID GLM-5.3-FlashX. The company's own numbers boil down to one figure: a peak output speed of 200 tokens per second, which Zhipu says is five times faster than the existing GLM-5.3-Flash.
The capability tier itself hasn't moved. According to the open-platform docs, Flash and FlashX share the same spec sheet: a 1-million-token context window, a maximum output of 128,000 tokens, input support for video, images, text and files, text-only output, plus thinking mode, tool calling, streaming responses, context caching and structured JSON output. The difference is stated plainly — FlashX trades up for throughput, while Flash holds the line on price.
Price climbs along with speed
Public reporting puts the pricing at 2 yuan per million input tokens and 7 yuan per million output tokens — about 2.5 times the original Flash tier. Zhipu didn't dwell on this in its own announcement, but on social platforms, the discussion has centered almost entirely on that speed-for-cost trade.
Working backward from the 2.5x multiple, Flash-tier output pricing comes out to roughly 2.8 yuan per million tokens. For an agent business burning through a billion output tokens a day, staying on Flash costs around 84,000 yuan a month on the output side; moving to FlashX pushes that to about 210,000 yuan. The extra 126,000 yuan buys the same volume of work done faster. Whether that's worth it comes down to whether the business is bottlenecked by first-token latency and queue depth. For conversational products, live captioning, or code completion — anything that can't tolerate a second's delay — 200 tokens per second and roughly forty-odd tokens per second are two different user experiences. For offline batch processing, document parsing, or overnight jobs where a few extra minutes cost nothing, Flash remains the better fit.
Where the 200 tokens/s comes from
Zhipu's explanation is that it layered further inference optimization on top of an inference-compute base built on 100,000 domestic chips. The statement isn't broken down any further — whether the optimization sits in speculative decoding, parallel scheduling, or memory orchestration goes unsaid, and whether that 100,000-chip figure covers Zhipu's own cluster alone or includes partner capacity is likewise undisclosed.
FlashX's predecessor has a track record abroad. GLM-5.3-Flash previously circulated among overseas developers under the code name Ox Alpha, with call volume that Zhipu once said exceeded DeepSeek's by more than double. From Ox Alpha to Flash to FlashX, Zhipu's playbook on this line has stayed consistent: use a low price to build up call volume first, learn the real shape of the load, then start charging by speed tier. When the official GLM-5.3 launched, million-token output was priced at $4.4; the Flash tier cut activated parameters down to 18 billion to compress cost; FlashX is another notch on that same curve, this time moving toward speed.
What the price hike signals
The default move among domestic large-model makers over the past year and change has been to cut prices — whoever cuts first grabs share. FlashX runs the other way: the same capability, a higher price tag, and a pitch built entirely around speed. That only works if inference-side supply has stopped being effectively free, and if concurrency headroom inside a cluster has itself become something that can be priced.
Whether it sticks comes down to two things: whether comparable domestic rivals follow with price increases of their own in the coming weeks, and whether FlashX can hold that 200-tokens-per-second ceiling during high-concurrency periods. Independent stress-test data isn't public yet.
Sources: Zhipu open platform documentation, Zhipu official announcement, CocoLoop, Sina Technology, Chinaz; speed and context specs verified against official docs, pricing per public reporting.