GLM-5.3-Flash Released

• Z.ai released GLM-5.3-Flash on 26 August 2026, weights and API together, twelve days after GLM-5.3; it is the first natively multimodal model of the GLM-5 series and takes image, video, text and file input^[1][4][5][6]
• The model arrived as a mystery: from 20 August an unattributed "Ox Alpha" was free on OpenRouter and OpenCode and topped OpenRouter's weekly usage table with 23.2 trillion tokens (to 25 August), twice DeepSeek V4 Flash 0731's 11.6 trillion, before Z.ai confirmed it was GLM-5.3-Flash^[4]
• Unlike GLM-5.3, which only re-trained GLM-5.2, Flash starts from a newly built base: 320B total and 18B active parameters, a 30-trillion-token multimodal corpus, hybrid sparse and linear attention and manifold-constrained hyper-connections; Z.ai says that cuts attention compute 3.01x and KV cache 4.44x against GLM-5.3^[1][6]
• Z.ai's headline is that it beats GLM-5.2 across its benchmarks at one-tenth of the price while approaching Claude Opus 4.8 on coding and agentic tests: Terminal Bench 2.1 84.3 against 85.0, DeepSWE v1.1 63.4 against 58.0 and AutomationBench 48.8 against 41.0^[5][6]
• Artificial Analysis scores it 57 on its Intelligence Index v4.1.1 at about $0.045 per task (at the launch discount), which put it on the cost/quality frontier in Z.ai's own chart^[4][5]
• Anthropic's 30 September red-team analysis later reported that Flash built a working Chrome exploit chain for roughly $20 of API fees plus eight hours of compute^[3]

Notable features

• 1M-token context window and 128K maximum output; video, image, text and file in, text out; thinking always on with low, high and max effort^[1][6]
• Listed at $0.15 per million input tokens, $0.03 cached and $0.50 output (half-price at launch on OpenRouter); GLM-5.3-FlashX costs $0.37, $0.075 and $1.25 and runs at about 200 tokens per second^[1][2][4]
• MIT-licensed weights with SGLang, vLLM, TokenSpeed and KTransformers deployment recipes^[5][6]
• Z.ai says the Ox Alpha period was served entirely on Chinese AI chips, on its own inference engine built over SGLang, with a 3x end-to-end serving gain over its first baseline on that hardware^[4]
• Vision is part of the coding loop: it turns screenshots or screen recordings into apps, drives browsers and desktop GUIs, builds Blender scenes and produces PPTX, PDF, DOCX and XLSX deliverables^[1]

Benchmarks

Z.ai's results chart for GLM-5.3-Flash (the model card carries the same chart as an image, not a table), transcribed from the chart on docs.z.ai.

AutomationBench v1.0.6; GDPval-AA v2 was run by Artificial Analysis. A dash means the chart shows no bar.