GLM-5.3 Flash

GLM-5.3 Flash is not a cheaper GLM-5.3. It is a different model, released August 26, 2026, and the first natively multimodal member of the GLM-5 family: text, image, and video in one stack rather than a vision head bolted onto a text model. Architecturally it is a 320 billion parameter mixture-of-experts activating roughly 18 billion per token across 45 layers, with hybrid linear and sparse attention (IndexPool and mHC) and a 30 trillion token multimodal pretrain. Z.ai cites roughly 3.0x less attention compute and a 4.4x smaller KV cache than text GLM-5.3. Context is 1 million tokens, weights are MIT licensed on Hugging Face at zai-org/GLM-5.3-Flash, and the model ran on OpenCode and OpenRouter under the stealth id ox-alpha before launch. Pricing is the headline and also the trap: list is $0.15 per million input tokens, $0.03 cached, and $0.50 output, roughly a tenth of the $1.40 input rate on text GLM-5.3, but a 50 percent launch promotion runs at $0.075/$0.25 through September 9, 2026 at 16:00 UTC. Budget the list rate. On the self-reported table the movement is concentrated in agentic work: DeepSWE 63.4 against GLM-5.2 at 46.2, AutomationBench 48.8 against 26.2, Toolathlon 78.4 against 59.9. Terminal Bench 2.1 at 84.3 is a 3.3 point gain and HLE with tools at 55.3 against 54.7 is a wash. On vision it leads DeepSeek V4 Flash Vision Exp on Chartography (78.0 to 64.3) and OfficeQA-Pro (62.4 to 57.9), and it loses BabyVision to Gemini 3.7 Flash 53.4 to 70.9. One operational constraint: thinking cannot be turned off, since thinking.type accepts enabled only.