GLM-5.3-Flash

Linas Beliūnas's Substack details the anonymous "Ox Alpha" model that appeared on OpenRouter and OpenCode on August 20, 2026, with a 1M-token context window and text, image, and video input but no disclosed creator. Within six days it became the most-used model on OpenRouter, processing roughly 23 trillion tokens, the On August 26, 2026, Bloomberg confirmed Ox Alpha as GLM-5.3-Flash from Zhipu AI, which operates internationally as Z.ai, with 320B total parameters and 18B active per token. It is the first natively multimodal model in the GLM-5 series and the cheapest model at its score on Artificial Analysis's Intelligence Index. MIT

Databricks confirmed on August 26, 2026 that Zhipu AI GLM 5.3 Flash is now available on Unity Gateway for multimodal reasoning, coding, and agentic workloads. Customers can access the model through the same governed API as other Unity Gateway models, gaining centralized permissions, guardrails, spend controls, and obse The same Unity Gateway release note also documents that the gateway can now track and cap spend on external model providers such as Amazon Bedrock and Azure AI Foundry via the External Model Spend in Budgets (Beta) feature. Customers are responsible for ensuring their compliance with applicable terms when using Zhipu A

Z.ai (formerly Zhipu AI) shipped GLM-5.3-Flash on August 26, 2026 under the tagline "Frontier Intelligence, Flash Cost," marking it as the first model in the GLM-5 series to be natively multimodal with image and video input alongside text. The model is a mixture-of-experts architecture with 320B total parameters and on Artificial Analysis's Intelligence Index v4.1.1 places GLM-5.3-Flash at 57 points and $0.045 per task, positioning it in the top-left "cheap and smart" corner of the cost-versus-intelligence plot that Z.ai claims delivers intelligence at roughly 10x lower cost than prior models. The article frames thinking mode as alwa

This piece from Greek Ai on Medium details Z.ai's August 26, 2026 release of GLM-5.3-Flash, an open-weight model under the MIT license with 320B total parameters and 18B active parameters. It is described as the newest member of the GLM-5 family designed around frontier-level capabilities at lower inference costs, comb The article covers the model's anonymous pre-launch as "Ox Alpha" on OpenRouter and OpenCode, where Z.ai gathered real-world developer feedback before the official reveal. It frames GLM-5.3-Flash as part of Z.ai (formerly Zhipu AI)'s strategy to deliver multimodal AI at dramatically lower inference costs. The piece ser

DataCamp reports that Z.ai released GLM-5.3-Flash as a cost-optimized sibling to its flagship GLM-5.3, which arrived under the anonymous preview codename "Ox Alpha" before public reveal. The model uses a 320B-A18B MoE design with a 1M token context and is natively multimodal. It scores 84.3 on Terminal-Bench 2.1, placi The blended price is roughly $0.10 per 1M tokens, about one-tenth of GLM-5.3's $0.90. Trade-offs include a 49 tokens/second output speed and a 306 GiB FP8 checkpoint that rules out lightweight local hosting. The article notes that Qwen3.8-Flash-Next launched the same day, intensifying competition in the budget-tier LLM

Ampere.sh provides a head-to-head comparison between GLM-5.3 and GLM-5.3 Flash based on Artificial Analysis testing. GLM-5.3 scores 60 on the Intelligence Index versus 57 for Flash, but Flash costs roughly one-ninth as much at normal API pricing and activates only 18B parameters versus 40B for GLM-5.3. GLM-5.3 Flash is A counterintuitive finding highlighted is that GLM-5.3 Flash is not actually faster at token generation, with Artificial Analysis measuring GLM-5.3 at ~85 tokens/second versus ~49 tok/s for Flash. Flash has a faster time-to-first-token at 1.46s versus 1.57s. The comparison recommends Flash for high-volume applications

A HuggingFace discussion thread documents a community PR adding evaluation results for zai-org/GLM-5.3-Flash extracted from the model's benchmark chart. The merged PR adds Terminal-Bench 2.1 (84.3), DeepSWE (63.4), and Humanity's Last Exam (55.3) to the model's evaluation results directory following the Hugging Face Hu The contributor notes that the model card cites arxiv:2602.15763 as "the GLM-5 Technical report," but that February 2026 paper is the original GLM-5 report and predates GLM-5.3-Flash, so it was not used as a source for specific numbers. The results were extracted visually from the model card's benchmark image since the

LLM-Stats ranks GLM-5.3-Flash at position 18 on its composite LLM Stats Score across tracked benchmarks. In capability sub-rankings, it places 3rd in Chat (of 117 models), 7th in Tool Calling (of 194), and 17th in Reasoning (of 363), while standing at 24th in Vision, 32nd in Coding, 38th in Math, 70th in Legal, and 86t Specific benchmark scores listed include GDPval-AA at 1773/3000 (rank 2), CharXiv-R at 0.89/1 (rank 8), and Terminal-Bench 2.1 results. The Quality Tracker shows improving performance with +0.89σ over 7 days and +8.43σ above baseline. Data sources for benchmarks are cited from Artificial Analysis, Z.ai, and the model's