A high-powered, unbranded artificial intelligence (AI) model dubbed Ox Alpha made a surprise debut on the API platform OpenRouter days ago, offering developers free access to a massive one-million-token context window and sparking intense industry speculation over its origins.
Appearing under the stealth identifier stealth/ox-alpha on Thursday, the mystery model features a 1.05-million-token context window, an expansive 131,072-token maximum output, and full multimodal capabilities supporting text, image, and video inputs. Promoted alongside open-source terminal agent OpenCode, the preview is billed as a zero-cost option for complex coding and long-horizon agentic tasks through a limited, week-long promotional window.
The launch features an ambitious capacity claim of up to 100 trillion tokens per day with zero data retention, offering a stark contrast to typical stealth releases that collect prompt data for training.
While the developer behind Ox Alpha remains strictly anonymous, early technical fingerprinting has led the AI community to point toward major Chinese technology labs. Developer analysis of the model’s tokenizer across multiple benchmark prompts revealed raw token counts nearly identical to Z.ai’s (formerly Zhipu AI) GLM-5.3, offset only by a standard wrapper. Other technical indicators, including specific “dirty token” errors, align with previous behaviors seen in the Qwen and GLM model families.
Alternatively, analysts point to Xiaomi’s MiMo team, which previously deployed stealth models on OpenRouter under pseudonyms like Hunter Alpha before official product announcements.
Despite claims circulating among early adopters of high scores on software engineering evaluations like DeepSWE, the anonymous model does not yet feature independent validation on formal public leaderboards. Industry experts urge caution, noting that social media benchmarks often fail to reflect long-term stability.
“Ox Alpha has enough verified capability on paper to deserve a serious trial,” noted one software analysis report. “It doesn’t have enough public evidence to deserve trust by default.”
The unannounced drop highlights a shifting landscape in global AI development.
According to platform metrics, the share of tokens processed by U.S.-based models on OpenRouter has plummeted from 70% to roughly 30% over the past year, with Chinese labs aggressively capturing market share through high-context, low-cost releases.
The arrival of Ox Alpha comes at a busy moment for OpenRouter, landing just days after corporate enterprise reports highlighted dramatic cost-savings from dynamic model routing, and as fintech giant Stripe Inc. moves forward with its acquisition of the router platform.
For now, developers can access stealth/ox-alpha directly through OpenRouter’s playground and API. However, security experts advise against inputting sensitive corporate codebases into any unverified model, regardless of zero-retention promises, until an official provider takes ownership.