Mysterious Ox Alpha beats Claude Fable 5 with 80% coding score

An anonymous model called stealth/ox-alpha has posted an 80% Pass@1 rate on the DeepSWE coding benchmark, beating Claude Fable 5's 65% and GPT-5.6-sol's 52%, and independent testing ties it to Zhipu's unreleased GLM flagship.

Ben Davis, the independent researcher who ran the tests, said he is "99% certain" the model belongs to Zhipu's GLM-5.x series, citing matching video-encoder behavior and tokenizer alignment.

The model appeared on OpenRouter on Aug. 20 with one week of free access, supports text, image and video input, and carries a 1.048 million-token context window. Across four controlled video sets, ox-alpha consumed tokens identically to GLM-5V-Turbo — frame-rate-independent sampling, roughly 147 tokens per second, and a per-frame resolution scaling mechanism — while candidates such as Xiaomi's MiMo v2.5 and Qwen 3.8 Max behaved differently. Token counts matched GLM-5.3 exactly across 25 prompts, with only a fixed +75 token wrapper difference.

If confirmed as Zhipu's unreleased multimodal flagship, the model would show a Chinese lab closing the coding gap with Anthropic and OpenAI, pressuring the valuation premium those companies command. Zhipu has not responded to requests for comment, and the free access window runs until Aug. 27.

Video Encoder and Tokenizer Point to Zhipu

Davis traced the model's origins across the video encoder, tokenizer, audio interface and output style. The video encoder proved the strongest signal: ox-alpha and GLM-5V-Turbo consumed video tokens in a completely identical manner across all four test sets, whereas MiMo v2.5, Qwen 3.8 Max and GLM-4.6V all showed distinct encoding characteristics. The model also rejects audio input, consistent with GLM-5V, while MiMo v2.5 — the leading competing candidate — supports audio. In output style, ox-alpha uses about 1.3 emojis per thousand characters, close to the GLM/Qwen series, while Claude, GPT-5.6 and Grok showed near-zero emoji usage in the same environment.

The report also ruled out other sources one by one. DeepSeek has not released video capabilities and uses a different tokenizer; Google, Qwen, xAI, OpenAI and Anthropic all mismatched on tokenizer, output style or video-encoder characteristics. Zhipu has a precedent for stealth testing — Pony Alpha was later confirmed to be related to GLM-5. GLM-5.3, the text-only model, shipped Aug. 14, and a unified vision flagship has been a focus of community attention.

Coding Gains Threaten the Frontier Lab Premium

On the capability side, ox-alpha passed eight of 10 deterministic DeepSWE tasks, including the "meriyah-explicit-resource-declarations" task on its first attempt, where GLM-5.3, GPT-5.6-sol and Grok 4.6 had all finished 0-for-4. It also cleared all 51,469 regression tests and, in an agent task spanning 69 tool calls, made a single error with no retry loops. Davis concluded the model is "significantly stronger" than GLM-5.3, suggesting a newer checkpoint rather than a variant.

The model's decoding speed differs from GLM-5V-Turbo by about 6 percent. GLM-5V-Turbo carries 744 billion total parameters with 40 billion active, leading Davis to speculate ox-alpha uses a Mixture-of-Experts architecture of similar scale — which would make the operator's claimed daily serving capacity of 100 trillion tokens more plausible. Sample sizes remain small, and the results need independent verification before they reshape the competitive picture.

For investors, the stakes are the premium the market assigns to frontier labs. Anthropic and OpenAI-related entities trade on the assumption their models lead on coding and reasoning; a Chinese challenger posting an 80% pass rate on DeepSWE, if verified, narrows that edge. Zhipu's stealth test on OpenRouter may serve as a public dress rehearsal ahead of an official launch, and past stealth models were claimed by their labs after free testing ended.

This article is for informational purposes only and does not constitute investment advice.