Z.ai, the Chinese AI company also known as Zhipu, confirmed on August 26 that it built Ox Alpha, the anonymous model that topped OpenRouter's leaderboard and set a single-day usage record after launching with no attribution and zero cost on August 201,2,5. The company simultaneously unmasked the model as GLM-5.3-Flash and released its weights on Hugging Face under the MIT license7.
GLM-5.3-Flash is a 320 billion-parameter hybrid model with 18 billion active parameters, trained specifically for ultra-low-cost inference9. Its architecture combines linear and sparse attention, reducing compute by 3x and KV-cache size by 4.4x compared with the full GLM-5.3 model. On OpenRouter, it is priced at $0.075 per million input tokens and $0.25 per million output tokens, with those figures reflecting a 50% discount.
Z.ai describes Ox Alpha as "a reasoning model designed for coding, sustained agentic work, and production workloads. It is suited for long-horizon software engineering, complex reasoning, and workflows that combine text with visual context". The model accepts multimodal text, image, and video input, supports native tool calling, function calling, and structured JSON output, and offers a 1,048,600-token context window with up to 131,000 tokens of output per call4.
Benchmark results drew attention but also dispute. On a small sample of ten DeepSWE software-engineering tasks, one independent researcher found Ox Alpha passed eight (80%), compared with 65% for Claude Fable 5 and 52% for GPT-5.6 Sol. On the full 113-task DeepSWE evaluation, Ox Alpha scored 63%. DeepSWE's own team re-evaluated and placed the model's overall capability around the level of Claude Opus 4.8 rather than clearly above Claude Fable 5. On the Artificial Analysis Intelligence Index, GLM-5.3-Flash scored 57 points, in line with GPT-5.6 Terra, Google Gemini 3.7 Flash, Meta Muse Spark 1.2, and Qwen 3.8 2.4T A95B.
Stripe CEO Patrick Collison called the model "very impressive," and the founder of HermesAgent said it beat several of the team's internal evaluation criteria. Cumulative API calls exceeded 4 trillion tokens13, and on OpenCode, Ox Alpha briefly ended DeepSeek's streak of 56 consecutive days at the top of the ranking.
Before the reveal, developers used technical fingerprinting to narrow the creator's identity: a tokenizer test found a constant 75-token difference from Zhipu's GLM-5.3, a video-encoder test matched GLM-5V-Turbo token-by-token, and an API error format echoed Zhipu's services. The stealth launch followed a pattern: earlier anonymous OpenRouter entries Pony Alpha and Hunter Alpha turned out to be Zhipu's GLM-5 and Xiaomi's MiMo-V2-Pro, respectively.
Z.ai also disclosed infrastructure details. The company said it achieved a 3x improvement in end-to-end serving performance on Chinese AI chips compared with its initial baseline on the same hardware, reaching per-token cost comparable to mainstream Nvidia GPUs. Z.ai built its own inference engine based on SGLang for the serving stack and said GLM-5.3 powered an infrastructure agent that created a feedback loop to optimize the system serving the model itself.
ANALYSIS The combination of MIT-licensed weights, aggressive pricing, and benchmark results competitive with Claude Opus 4.8 and GPT-5.6 Terra positions GLM-5.3-Flash as a direct cost competitor to Western frontier models in agentic and coding workloads. The Chinese-chip serving claims, if reproducible, carry additional weight given ongoing export restrictions on Nvidia hardware.