Ox Alpha (Fully Tested) : So, This is GLM-5.5?! IT'S CRAZY!
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A new anonymous stealth model called Ox Alpha appeared on Open Code, offering a free trial period, a 1 million token context window, multimodal support, and zero data retention, with claimed capacity of 100 trillion tokens per day. On the creator's Kingbench, it scored 87.5%, placing second behind GLM 5.3, and outperforming models like Opus 4.8 and Qwen 3.8 Max. On a separate 10-task Deep SWA subset run by Ben Davis, it scored 80%, beating GLM 5.3, Grok 4.6, and GPT 5.6. Fingerprinting analysis (video encoder token counts, tokenizer vocabulary matches, writing style, audio rejection) points with about 90% confidence to this being Zhipu's (GLM maker) next-generation unified multimodal model, though it remains unconfirmed. The free period is expected to end around August 27th, when the model's identity may be revealed.
Questions this post answers
What is the Ox Alpha stealth model and which company is likely behind it?
Ox Alpha is an anonymous stealth model that appeared on Open Code with a 1 million token context window, multimodal support, and zero data retention, offered free for a limited period with claimed capacity of 100 trillion tokens per day. Fingerprinting analysis of video encoder token counts and tokenizer vocabulary matching GLM 5.3 and GLM 5V Turbo suggests around 90% confidence it is Zhipu's next-generation unified multimodal GLM model, though this remains unconfirmed. Developers tracking which lab actually ships this model can follow the reveal coverage on daily.dev.
How does Ox Alpha perform compared to GLM 5.3, Opus 4.8, and GPT 5.6 on coding benchmarks?
Ox Alpha scored 87.5% on Kingbench, placing second behind GLM 5.3's 91.25% but ahead of Opus 4.8 at 80% and Qwen 3.8 Max at 81.25%. On a separate 10-task Deep SWA subset, it scored around 80%, beating GLM 5.3 (62%), Grok 4.6 (62%), and GPT 5.6 (52%), and solved a Marriott task in one shot where those three models scored zero. Teams comparing agentic coding models can weigh these benchmark gaps before picking a default on daily.dev.