Ox Alpha Tops OpenRouter as Top Agent Apps Push Usage Past 4 Trillion Tokens

An anonymous AI model called Ox Alpha has rapidly become one of the most heavily used models on OpenRouter.

The model appeared on OpenRouter on August 20, 2026 under the identifier stealth/ox-alpha. In the Chinese AI community, it has also picked up the nickname “Niu Lai.”

AIBase reported on August 24 that Ox Alpha had quickly climbed to the top of OpenRouter’s model usage rankings and set a new single-day usage record. More importantly, its traffic was no longer coming only from curious users testing a mysterious new model.

Large agent and coding applications were already sending enormous amounts of production-style workload through it.

Ox Alpha Quickly Climbed to the Top of OpenRouter

According to the original AIBase report, Ox Alpha’s usage rose sharply shortly after it became available on OpenRouter.

That growth is notable because the model launched without a public developer identity, conventional product announcement, detailed technical report, or recognizable model family name.

OpenRouter currently describes Ox Alpha simply as a reasoning model designed for coding, sustained agentic work, and production workloads.

Its intended scenarios include:

• Long-horizon software engineering.
• Complex reasoning.
• Persistent agent workflows.
• Tasks combining text with visual context.

OpenRouter also makes clear that it is only routing requests to the model.

The platform says Ox Alpha is developed and operated by an anonymous third-party provider and that OpenRouter is not the developer, owner, or model provider.

Top Agent Applications Have Sent More Than 4 Trillion Tokens Through Ox Alpha

The biggest signal in the AIBase article is not a benchmark score. It is real usage.

At the time the source article was published, AIBase said the five largest applications using Ox Alpha—including Hermes Agent and Claude Code—had already accumulated more than:

4 trillion tokens

of Ox Alpha usage.

OpenRouter’s live counters continue to change rapidly, so this should be treated as a snapshot from the article’s publication time rather than a permanent total.

The current Ox Alpha model page shows major usage coming from applications such as:

This pattern supports the point made in the original article: Ox Alpha’s popularity is moving beyond community experimentation and into real agent workflows.

That matters because sustained software-development agents can consume far more tokens than ordinary chat sessions. A single task may involve reading large repositories, planning across many steps, calling tools, reviewing output, retrying failures, and continuing for long periods.

For this kind of model, token volume can therefore become a useful signal of whether developers are actually putting it to work.

What OpenRouter Officially Confirms About Ox Alpha

Although the developer remains anonymous, OpenRouter currently publishes several concrete specifications.

These specifications help explain why the model has attracted agent developers.

A context window above one million tokens gives applications much more room for large repositories, long task histories, tool outputs, and multimodal context.

The free preview also lowers the cost of running unusually token-intensive experiments.

OpenRouter warns, however, that stealth models are preview models and may only be available for a limited period.

The Developer Behind Ox Alpha Is Still Unknown

The second major theme in the AIBase article is the model’s identity.

As of publication, no company has publicly claimed Ox Alpha.

OpenRouter’s wording is explicit: the model is operated by a third-party provider that has chosen to remain anonymous during the preview.

That leaves room for speculation, but not confirmation.

Community Analysis Points Toward the GLM Family

The original article notes that community researchers have compared technical fingerprints such as tokenization behavior and found similarities with Z.ai’s GLM family.

Other independent investigations have reported similar serving-layer or tokenizer overlaps.

That evidence can be interesting, but it does not prove ownership.

Two models can share tokenization behavior, infrastructure, wrappers, or serving technology without being the same model or coming from the same organization.

Z.ai has not officially claimed Ox Alpha.

Gemini Speculation Also Appeared

AIBase also reports that several Google DeepMind researchers made social posts that some users interpreted as hints that Ox Alpha could be related to a new Gemini model.

Again, no official confirmation followed.

Google has not identified Ox Alpha as a Gemini model, and OpenRouter still lists the provider only as Stealth.

The safest description remains:

Ox Alpha’s developer is unknown. GLM, Gemini, and other identity theories remain speculation until a vendor publicly confirms the model.

Free Access Comes With an Important Data Consideration

OpenRouter currently lists Ox Alpha at zero cost for both input and output tokens.

Its model page also says prompts and completions are retained by the anonymous provider and are not used for training under the model-specific notice shown there.

OpenRouter’s broader Stealth Program terms contain additional provisions governing collection, sharing, retention, evaluation, and improvement of user content.

Because the provider is anonymous and preview terms can change, developers should review the live model page and Stealth Program terms before sending:

• Proprietary source code.
• Customer data.
• Credentials.
• Internal documents.
• Other sensitive information.

The model may be attractive for large-scale experimentation, but anonymous-provider preview access is different from a normal enterprise relationship with a known model vendor.

Real Workflow Usage Is Becoming a More Important Model Signal

The final point in the source article is broader but closely tied to Ox Alpha’s rise.

AI models are increasingly moving from isolated benchmark tests into real agent workflows.

In an agent environment, the model may be asked to:

• Read a codebase.
• Modify files.
• Run tools.
• Interpret errors.
• Plan multiple steps.
• Continue working over long sessions.

That makes real application usage increasingly informative.

A benchmark can tell developers whether a model performs well under a controlled evaluation.

Usage data can show whether developers continue choosing the model once it is connected to actual tools and workloads.

Ox Alpha’s rapid token growth therefore does not prove that it is the best model on every benchmark.

It does show that developers and agent platforms have been willing to send very large workloads through it.

For an anonymous preview model, that level of adoption is itself unusual.

FAQ

What is Ox Alpha?

Ox Alpha is an anonymous reasoning model available through OpenRouter under the model ID stealth/ox-alpha. OpenRouter positions it for coding, long-horizon software engineering, sustained agentic work, complex reasoning, and multimodal workflows.

Who made Ox Alpha?

The developer has not been publicly identified. OpenRouter says the model is developed and operated by an anonymous third-party provider and that OpenRouter itself is only routing requests to it.

Did Ox Alpha really exceed 4 trillion tokens?

AIBase reported that, at the time of publication, the five largest application routes using Ox Alpha had collectively exceeded 4 trillion tokens. OpenRouter’s usage counters are live and continue changing, so the figure should be treated as a dated usage snapshot.

Is Ox Alpha free?

OpenRouter currently lists both input and output pricing at zero. Because Ox Alpha is a stealth preview model, availability and pricing may change or the model may be removed with limited notice.

How large is the Ox Alpha context window?

OpenRouter lists a context window of 1,048,576 tokens and a maximum completion length of 131,072 tokens. It accepts text, images, and video as input and returns text.

Is Ox Alpha actually a GLM model?

That has not been confirmed. Community researchers have reported technical similarities with the GLM family, but shared tokenizer or serving fingerprints are not enough to establish the model’s developer.

Is Ox Alpha a new Gemini model?

There is no official confirmation from Google or Google DeepMind. The Gemini theory remains community speculation, while OpenRouter continues to list the provider as anonymous.

Is Ox Alpha safe for proprietary code?

Developers should review the current OpenRouter model-specific notice and Stealth Program terms before using sensitive material. The provider is anonymous, and OpenRouter says prompts and completions are retained by that provider under the conditions shown on the live model page.

Related Tools

• OpenRouter: Multi-model API platform currently hosting Ox Alpha under the stealth/ox-alpha route.
• OpenCode: Agentic coding platform with model usage data and Ox Alpha availability.
• Hermes Agent: Nous Research’s open-source persistent AI agent and one of the largest Ox Alpha application routes shown by OpenRouter.
• Claude Code: Anthropic’s agentic coding tool, also listed among major applications sending traffic to Ox Alpha through OpenRouter.

Related Links

• Ox Alpha on OpenRouter: Official live model page with provider status, context window, pricing, modalities, tool support, and application usage.
• OpenRouter AI Model Rankings: Live rankings based on real-world token usage across OpenRouter.
• OpenRouter Stealth Program Terms: Terms governing anonymous preview models and handling of user content.
• OpenRouter API Quickstart: Official developer documentation for accessing OpenRouter models through its API and SDKs.
• Ox Alpha on OpenCode Data: OpenCode’s usage page for the anonymous Ox Alpha model.
• Hermes Agent on OpenRouter: OpenRouter usage page for one of Ox Alpha’s largest application routes.
• Claude Code on OpenRouter: OpenRouter usage page for Claude Code and its model traffic.

Summary

Ox Alpha appeared on OpenRouter as an anonymous stealth model on August 20 and quickly attracted unusually large traffic. AIBase reported that its five largest application routes had already exceeded 4 trillion tokens of cumulative usage, with agent and coding tools such as Hermes Agent and Claude Code contributing heavily.

OpenRouter confirms that Ox Alpha is a free preview reasoning model with a 1,048,576-token context window, multimodal input, tool calling, and a focus on coding and sustained agent workflows. What OpenRouter does not reveal is the developer.

Community evidence has fueled theories involving the GLM family and Gemini, but no vendor has officially claimed the model. Until that changes, those identity theories should remain clearly labeled as speculation.

The strongest verified signal around Ox Alpha is not its mystery identity—it is the speed at which real agent applications began putting trillions of tokens through it.