A new artificial intelligence model known as Ox Alpha has rapidly attracted attention from developers after appearing without any named creator or formal announcement. Listed on major routing platforms as a stealth release from a third-party provider that elected to stay anonymous during its preview period, the system is positioned for coding tasks, extended agent workflows and production use.

Available since approximately August 20, the model supports text, image and video inputs while generating text outputs. It features a claimed context window of 1,048,576 tokens and a maximum output of 131,072 tokens. Tool and function calling are supported, though structured JSON responses are not schema-enforced in all cases.

Free Access Fuels Rapid Adoption

During the limited preview, Ox Alpha is offered at no cost for both input and output tokens through platforms including OpenCode Zen, the paid OpenCode Go subscription, and other routing services. OpenCode described near-unlimited rate limits and capacity to process 100 trillion tokens per day. Separate claims from other providers suggested even higher throughput figures.

Community reports indicate substantial early usage. In its first week, the model reportedly handled roughly 7.1 trillion tokens across about 134,000 users. Later tracking on one platform showed tens of trillions of tokens processed and hundreds of thousands of unique users within days of launch. The free window is expected to last approximately one week from debut, after which pricing details remain undisclosed.

Stripe chief executive Patrick Collison tested the system and described it as very impressive, adding momentum to early discussion. The combination of zero cost, large context capacity and multimodal support has encouraged extensive real-world testing, particularly among developers running long agent sessions that would otherwise incur significant expense.

Performance Signals and Benchmark Claims

Early independent tests have generated notable results on coding benchmarks. One developer ran a 10-task subset of DeepSWE, which measures successful resolution of real GitHub issues on the first attempt. Ox Alpha scored 80 percent, ahead of Claude Fable 5 at 65 percent and GPT-5.6 Sol at 52 percent. Full runs across the complete 113-task set later placed the model near 63 percent, roughly on par with GPT-5.6 Sol.

As of late last week the model lacked official entries on major independent evaluation platforms. Capacity figures and performance claims have not undergone broad third-party verification. The system is described as a reasoning model suited to long-horizon software engineering, complex problem solving and workflows that combine textual and visual context.

Intense Speculation Over the Creator

No organization has claimed responsibility for Ox Alpha. The listing states only that it comes from a third-party provider remaining anonymous for the preview. Developers and analysts have examined tokenizer behavior, error patterns, video token accounting and other technical fingerprints in an effort to identify the source.

The leading theory centers on Zhipu AI, also known as Z.ai, and its GLM model family. Multiple independent probes found close matches with GLM-5.3 tokenizer characteristics and video-token usage patterns. The company has previously released models under anonymous codenames, including an earlier test of GLM-5 as Pony Alpha. Zhipu has made no public comment on the current release.

Other suggestions have included an unreleased Microsoft MAI variant, Google systems, Xiaomi models and additional Chinese or U.S. laboratories. Many of these alternatives were weakened by mismatches in input modalities, tokenizer details or release practices. One analyst noted that initial confidence in specific theories faded quickly as more data emerged, leaving the origin unresolved.

The Broader Pattern of Stealth Releases

Anonymous model launches have become a recurring practice. Labs use them to gather genuine usage data and feedback before attaching a brand name, reducing the risk of an underwhelming public debut. Prior examples include systems later linked to Google, OpenAI checkpoints, Xiaomi and Meituan. Several other codenamed models have appeared and never been formally attributed.

In the current case, free high-capacity access creates strong incentives for testing while generating organic discussion and additional data for the unknown provider. The approach allows collection of real-world performance signals on coding and agent tasks at relatively low marketing cost. Whether Ox Alpha follows the pattern of previous stealth models and receives an official identity in the coming days remains to be determined.

Developers continue to evaluate the system across available routes while the preview remains free. Technical characteristics and early usage volumes have established it as one of the more closely watched AI releases of the month, even as its creator stays silent.