Mystery Model 'Ox Alpha' Debuts Free on OpenRouter, Outperforms Claude — Clues Point to Zhipu

Forum topic · QianXun · 2026-08-23

Summary

OpenRouter quietly listed an anonymous AI model called Ox Alpha on August 20, free for one week, with its developer unknown. Developer Ben Davis benchmarked it on DeepSWE — a test of real-world coding ability involving navigating codebases, finding bugs, and writing working patches — where Ox Alpha scored roughly 80% across ten tasks, beating Claude Fable 5 (65%), GLM-5.3 (62%), Grok 4.6 (62%), and GPT-5.6 Sol (52%). OpenRouter specs list a 1,048,576-token context window, multimodal input (text, image, video), up to 131,072 output tokens, and free access during the preview period; OpenCode also integrated it into its free tier, claiming up to 100 trillion tokens processed per day. Tokenizer fingerprinting across 25 prompts shows Ox Alpha's token counts nearly identical to GLM-5.3, offset by a constant 75-token wrapper, and its video encoder behavior and API response characteristics match Zhipu's architecture — evidence pointing to Zhipu (Z.ai) as the likely developer.

The world's largest AI model aggregator, OpenRouter, quietly listed an anonymous model on August 20 called Ox Alpha, offered free for one week. Nobody knows which company is behind it.

Benchmark Performance

Developer Ben Davis ran it through the DeepSWE benchmark — which measures an AI's ability to read real codebases, find vulnerabilities, and write usable patches. Across ten tasks, Ox Alpha scored about 80%. By comparison, Claude Fable 5 scored 65% and GPT-5.6 Sol only 52%. Davis himself admitted the result left him puzzled, since the model's origin is completely unknown.

| Model | Average score on 10 DeepSWE tasks | |-------|-----------------------------------| | Ox Alpha (anonymous) | 80% | | Claude Fable 5 (max) | 65% | | GLM-5.3 (max) | 62% | | Grok 4.6 (xhigh) | 62% | | GPT-5.6-sol (max) | 52% |

Specs and Availability

OpenRouter's listed specifications:

Context window: 1,048,576 tokens
Input: text, images, and video
Max output: 131,072 tokens per response
Pricing: free during the preview period

OpenCode, another company building a terminal coding agent, has also integrated it into its free tier, claiming it can process up to 100 trillion tokens per day.

Is It Zhipu?

So is it Zhipu? A tokenizer fingerprint analysis across 25 prompts found that Ox Alpha's token counts are almost identical to GLM-5.3's, differing only by a constant 75-token wrapper. The video encoder's behavior and the API's response characteristics also match Zhipu's architecture. For now, all the clues point to the same company.

*(Adapted from an IT之家 report dated 2026/8/22: "Mystery Ox Alpha AI model launches with free limited access: DeepSWE scores crush Claude Fable 5, clues point to Zhipu.")*