{"slug": "z-ai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek", "title": "Z.AI Reveals It Secretly Built Ox Alpha, the Model That Beat DeepSeek", "summary": "Z.AI confirmed it built Ox Alpha, the anonymous reasoning model that topped OpenRouter's leaderboard on August 20 and surpassed DeepSeek, releasing the open weights for free. The model, known informally as GLM-5.3 Flash, is the newest version of its GLM series and is priced at $1.40 per million input tokens and $4.40 per million output tokens under an MIT license. The stealth launch generated a week of uncredited buzz, including a 'very impressive' endorsement from Stripe CEO Patrick Collison, and comes ahead of Z.AI's first detailed earnings report since its January 2026 IPO.", "body_md": "*Z.AI has confirmed it built Ox Alpha, the anonymous model that stormed OpenRouter's leaderboard last week and knocked DeepSeek out of the top spot. The company is releasing the weights for free.*\n\nFor six days, nobody would say who made it. Ox Alpha showed up on OpenRouter on August 20 under the provider name \"stealth,\" free to use, with a context window just over a million tokens and no name attached beyond a marketing description: a reasoning model built for coding, sustained agentic work, and production workloads. By Wednesday, Z.AI ended the guessing game. The Chinese lab told Bloomberg that Ox Alpha is the newest version of its GLM series, and it released the open weights that night.\n\nIn between, Ox Alpha didn't just get noticed. It took over. The model surged past DeepSeek, which had held the top of OpenRouter's leaderboard for 56 straight days, and according to Bloomberg its usage on the platform more than doubled DeepSeek's. Coding tools including Claude Code pushed billions of tokens through the model within days, Bloomberg reported, while OpenCode advertised capacity of up to 100 trillion tokens a day for anyone routing traffic through it.\n\nStripe CEO Patrick Collison called it \"very impressive\" on X, and that sent the internet hunting for a name. Some guessed Microsoft's unreleased MAI models. Most bet on China. The bet paid off through unglamorous detective work: developers ran tokenizer probes, comparing how Ox Alpha broke text into tokens against known models. On one widely cited test, Ox Alpha matched GLM on all 11 probes run against it. No other model came close to that many matches. DeepSeek, for comparison, used 98 tokens on a digit-heavy test where GLM needed just 29.\n\nThe benchmark numbers backed up the hype, mostly. An early, informal 10-task run on the DeepSWE coding benchmark put Ox Alpha at 80% first-pass accuracy, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. A larger test run afterward brought Ox Alpha back down closer to GPT-5.6-sol's level. Impressive. Just not the runaway win the first numbers suggested.\n\n[Chinese State Hackers Doubled Their Attack Volume Using DeepSeek AI](https://startupfortune.com/chinese-state-hackers-doubled-their-attack-volume-using-deepseek-ai/)\n\nChinese state-affiliated hacking groups have more than doubled their attack volume after integrating DeepSeek into their operations, according to Taiwanese research firm TeamT5. Separately, Palo Alto Networks' Unit 42 found an independent hacker used DeepSeek to autonomously scan and breach 14 of more than 460 targeted systems. - [chinese state hackers using deepseek AI for attacks](https://startupfortune.com/chinese-state-hackers-doubled-their-attack-volume-using-deepseek-ai/) - [how AI models are scaling cyberattack operations](https://startupfortune.com/chinese-state-hackers-doubled-their-attack-volume-using-deepseek-ai/)\n\n## Why a lab would hide its own launch\n\nHere's the thing about stealth releases: they work. Z.AI got a full week of uncredited buzz - screenshots, benchmark threads, a Stripe CEO's endorsement - before ever putting its name on the model. That's a week's head start. Once the mystery broke, it read less like a leak and more like a plan. Z.AI's GLM-5.3, released earlier in August, already claims to rival Anthropic's Fable 5 on some benchmarks. Ox Alpha, known informally as GLM-5.3 Flash, is the free, faster, more aggressively marketed sibling.\n\nThe economics tell the rest of the story. GLM-5.3 prices out at $1.40 per million input tokens and $4.40 per million output tokens, with cached input running about 81% cheaper still, and it ships under an MIT license that lets anyone build on it commercially. That's a fraction of what OpenAI or Anthropic charge for comparable reasoning models.\n\nThere's a shareholder angle here too. Z.AI went public in January 2026, and Bloomberg notes the company's first detailed earnings report since listing lands Monday. A viral, uncredited launch that gets stamped with the company's name right before that report is a tidy story to walk into an earnings call with.\n\nThis isn't an isolated data point. According to Techstrong.ai, the share of tokens processed by U.S.-based models on OpenRouter has fallen from roughly 70% to about 30% over the past year, as Chinese labs including DeepSeek, Alibaba's Qwen and Z.AI's GLM line push in with cheaper, openly licensed alternatives. DeepSeek's 56-day reign at the top of that leaderboard wasn't a fluke. It was the start of a pattern Z.AI just extended.\n\nWhat OpenAI and Anthropic do next is the open question. Neither has matched Chinese labs on price. Neither gives away model weights. Z.AI says a full model card, with exact parameter counts, is coming soon.\n\n**Also read:** [Legato Emerges From Stealth With $12M to Put AI Hearing Aids Into Glasses](https://startupfortune.com/legato-emerges-from-stealth-with-12m-to-put-ai-hearing-aids-into-glasses/) • [Deutsche Bank Lets Google's Gemini AI Help Judge Credit Risk](https://startupfortune.com/deutsche-bank-lets-googles-gemini-ai-help-judge-credit-risk/) • [Perplexity's Portable Computer Runs AI Agents Free on Your Own Nvidia GPU](https://startupfortune.com/perplexitys-portable-computer-runs-ai-agents-free-on-your-own-nvidia-gpu/)", "url": "https://wpnews.pro/news/z-ai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek", "canonical_source": "https://startupfortune.com/zai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek/", "published_at": "2026-08-26 15:35:52+00:00", "updated_at": "2026-08-26 15:45:33.917918+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "generative-ai", "ai-products", "ai-research"], "entities": ["Z.AI", "Ox Alpha", "DeepSeek", "OpenRouter", "GLM-5.3 Flash", "Stripe", "Patrick Collison", "Bloomberg"], "alternates": {"html": "https://wpnews.pro/news/z-ai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek", "markdown": "https://wpnews.pro/news/z-ai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek.md", "text": "https://wpnews.pro/news/z-ai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek.txt", "jsonld": "https://wpnews.pro/news/z-ai-reveals-it-secretly-built-ox-alpha-the-model-that-beat-deepseek.jsonld"}}