A model from a developer listed only as “Stealth” appeared quietly on the LLM marketplace OpenRouter. Its description read simply: built for coding, long-horizon agent tasks and production. For its first week it offered free capacity of up to 100 trillion tokens a day. Called Ox Alpha, it jumped to first place on OpenRouter’s leaderboard by a wide margin.
Within 48 hours the industry was paying attention. One X user ran 10 tasks from the DeepSWE benchmark: Ox Alpha passed eight, while Claude Fable 5 scored 65% and GPT-5.6 Sol 52% under the same conditions. The results were unofficial, but reaction was explosive. Ox Alpha processed 27 trillion tokens in a single week, more than double second-place DeepSeek V4 Flash at 12 trillion.
Researchers found its tokenizer and error patterns closely matched Zhipu AI’s (Z.ai) GLM family. On August 26, Z.ai confirmed to Bloomberg that it was a preview of its new model, GLM-5.3-Flash — which became paid after official launch.
Three motives are widely suspected:
Nvidia released Nemotron 3.5 Lightning for free in August — a lightweight open-source model that runs on a single GPU and allows commercial use, modification and redistribution. As Jensen Huang has framed it, free AI grows GPU demand: more companies adopt AI, then buy GPUs to run it. Built for fast, repetitive work, the model is designed as an agent’s “hands and feet.” With examples like Kimi k3, low-cost and free models with near-frontier performance keep arriving.
For users, better models are getting cheaper. As AI agents take on more tasks, choosing and combining the best model for each job — a multi-model strategy — becomes more important than standardizing on one. The source cites a figure that 79% of companies have rolled out agentic AI but only 11% use it in real work, pointing to integration methods as the challenge for the second half of 2026.
For marketing teams:
As model performance converges, differentiation shifts to narrative and operations — see the story encoded in Nvidia’s Hugging Face price and the quiet launch of Claude Fable 5.1.
A preview of Zhipu AI’s (Z.ai) GLM-5.3-Flash, confirmed via Bloomberg on August 26. It became a paid model after official launch.
Likely for blind feedback, large-scale infrastructure testing and low-cost buzz. For hardware makers like Nvidia, free models also drive GPU demand.
Choosing and combining the best model for each task based on quality and cost, rather than using a single model for everything, validated on your own work samples.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Buzzvil's BD strategy team has zero developers but runs 32 AI workflow automations — from 7-minute revenue drop analysis to 90% fewer manual steps. How they scaled from personal hacks to team systems.
Meta’s Muse Wonder and Muse Wonder Full add Active Memory and Active Focus so its AI agent can handle tasks that take days. But a Toronto user’s Marketplace mishap shows the risks of agentic AI.
Pew cloned thousands of panelists as AI digital twins and found a 12.4-point average error. Five ways synthetic respondents fail — and where marketers can still use them in research.