An anonymous provider has been handing out free access to a frontier-class coding model on OpenRouter Inc. since Aug. 20, and no company has admitted to building it.

The model is listed as “Ox Alpha” and costs nothing for input or output tokens. OpenCode, an open-source terminal agent, debuted the model on the same day and advertised capacity of 100 trillion tokens a day. Rate limits there were described as near unlimited.

OpenRouter told users in an email that free access on its route runs through Monday, Aug. 24. OpenCode’s launch note promised a week from Aug. 20, which points to about Aug. 27.

The model’s context window runs to 1,048,576 tokens and output caps at 131,072 tokens. The model takes text, images and video. Audio requests come back rejected. OpenRouter’s listing pitches it as a reasoning system for long-horizon software engineering and sustained agentic work.

Early scores got the attention. Ben Davis, a developer posting as @davis7, ran the model through 10 tasks on the DeepSWE software engineering benchmark and scored it above 80%. Claude Fable 5 managed 65% on the same tasks. GPT-5.6-sol got 52%. Davis flagged the sample size himself, warning that a subset that small carries a lot of variance.

The hedge was warranted. Davis has since run the full DeepSWE set, and Ox Alpha came out roughly level with GPT-5.6-sol mid rather than well ahead of it. No public leaderboard has posted a result.

The work on identifying the lab has been more rigorous than the scoring. A developer known as unclecode, who wrote the Crawl4AI open-source crawler, built a browser tool called modelprint that fingerprints anonymous API endpoints. Nine infrastructure probes go out to an endpoint and the responses get matched against known models. Ox Alpha lined up with GLM-5.3 on six of the nine. That included all four normalized tokenizer counts. No other lab’s best candidate cleared more than two of the four.

He is careful about what that proves. “Matching fingerprints prove shared infrastructure, not identity,” unclecode wrote in the project documentation. A lab can serve two different models on the same stack, he noted.

Chinese artificial intelligence developer Z.ai Co. has previewed anonymously before. The company ran GLM-5 on OpenRouter under the name Pony Alpha ahead of its release. GLM-5.3 shipped Aug. 14, six days before Ox Alpha turned up.

Other testers are not sold. Xiaomi Corp.’s MiMo team has come up more than once, and it has previewed unbranded models before. One reading of the tokenizer behavior points instead to cl100k_base. That encoding came out of OpenAI Group PBC and sits oddly on a Chinese model.

The terms have drawn as much attention as the benchmarks. OpenRouter’s model page says prompts and completions are retained by the provider and are not used for training. Broader terms covering anonymous previews on the site extend to training, evaluation and improvement. OpenCode’s route advertises zero retention from a provider it does not name.

That gap matters more if the Z.ai attribution holds, as the U.S. Commerce Department added Zhipu AI to its Entity List on Jan. 16, 2025, in a rule that said the listed companies advance China’s military modernization through the development and integration of advanced artificial intelligence research. Zhipu AI is Z.ai’s former name. Enterprise code is flowing to a provider with no name attached. Nobody sending it can check where it lands.

Usage has not slowed. Coding tools including Claude Code have pushed billions of tokens through the model since Thursday, Bloomberg reported Sunday.

OpenRouter says it’s not the developer, owner or operator of Ox Alpha and only routes requests to it.

Image: SiliconANGLE/GPT Image 2