Breaking OpenRouter's historical record! The mysterious bullish model Ox Alpha goes viral overseas.

According to media reports, the AI community both domestically and internationally has been stirred up recently by a mysterious model.

A model called Ox Alpha quietly appeared on OpenRouter and OpenCode, and on its first day shot to the top of OpenRouter's leaderboard, setting a new record for single-day model usage.

The OpenRouter page shows that cumulative usage from just the top five entry points, including Hermes and Claude Code, has already exceeded 4 trillion tokens.

In addition, Ox Alpha ended DeepSeek's 56-day streak at the top of OpenCode, becoming the new "model" that developers worldwide have been flocking to over the past two days.

According to information released by the platform, Ox Alpha is defined as a cutting-edge AI model focused on code writing, long-horizon agent tasks, and real-world production environments.

In terms of core parameters and technical specs, it demonstrates an extremely high ceiling: its context window capacity reaches approximately 1.048 million tokens, with a maximum output length of 131,000 tokens, and it fully supports multimodal input across text, image, and video. During the preview period, it is being offered to developers as a free API.

Meanwhile, the open-source terminal coding agent OpenCode also announced integration of the model on the same day, preparing a staggering daily call capacity of up to 100 trillion tokens.

Since "Ox" literally means "bull" in English, and combined with the recent viral meme from the abstract film "The Bull Comes," Chinese netizens quickly gave it a fitting nickname—"Bull Comes LLM."

Within just hours of its release, Ox Alpha attracted intensive benchmarking from a large number of developers. Independent researcher Ben Davis subjected it to rigorous testing using 10 subtasks from the software engineering benchmark DeepSWE.

The results showed that Ox Alpha achieved a pass rate of approximately 80%, outperforming a range of well-known industry models, including Claude Fable 5 at 65%, GLM-5.3 and Grok 4.6 at 62%, and GPT-5.6 Sol at 52%.

However, industry insiders also noted that this data currently comes from preliminary testing by an individual researcher, and the official DeepSWE leaderboard has not yet formally included the model.

Several Google DeepMind researchers posted hints suggesting this is a new version of Gemini, fueling rapid speculation linking Ox Alpha to Google's new model.

But the community, using tokenizer fingerprints and error codes as evidence, has pointed toward Chinese model maker Zhipu AI's GLM series. Zhipu has not yet confirmed ownership of the model.

Notably, OpenCode has claimed it has prepared a daily service capacity of 100 trillion tokens for Ox Alpha, allowing developers worldwide to conduct near-unlimited stress testing in real-world scenarios.

If the model's Chinese identity is later confirmed, it would mean that domestic large models are not only approaching the frontier level of overseas models in benchmark scores, but may also already possess the capacity to handle trillion-token calls and support global agent workloads at scale.