Summary
- Ox Alpha, a free "stealth model," was launched on August 20 via OpenRouter, OpenCode, Cline, and Nous Research's portal.
- Initial reports claim Ox Alpha surpasses Claude Fable 5 and GPT-5.6 Sol based on a sample of 10 tasks; however, two separate extensive evaluations of 113 tasks yielded results close to 63%, aligning with GPT-5.6 Sol.
- Speculations about the model's origins include Google, Xiaomi, DeepSeek, and Microsoft's MAI, but the prevailing theory attributes it to Zhipu AI's GLM-5.3 series.
The AI community is buzzing about a newly unveiled model named Ox Alpha, which debuted on OpenRouter without any identifying company details on August 20.
Described as "a reasoning model optimized for coding, continuous agentic tasks, and production workloads," the model has been developed by an unnamed third-party provider during its preview phase.
Notably, this enigmatic AI model has outperformed Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol in several coding benchmarks.
The DeepSWE benchmark measures the success rate of a coding agent in accurately resolving actual GitHub issues on the first attempt, with results calculated from 113 tasks across 91 repositories. Ox Alpha shows competitive results against top-tier models.
Developer Ben Davis conducted an initial test with 10 tasks, achieving an 80% success rate for Ox Alpha on the first attempt, compared to 65% for Claude Fable 5 and 52% for GPT-5.6 Sol. This statistic quickly gained traction online, leading to widespread speculation that Ox Alpha holds a definitive advantage over its rivals.
I ran this thing through 10 tasks on DeepSWE (so there could be a ton of variance in its real score, this is a subset), but uh...
gpt-5.6-sol: 52%
fable: 65%
whatever the hell this is: 80% (was a near miss on the "x"s so actually over 80%)I am very confused https://t.co/VfH6n5i7xc pic.twitter.com/NdDSTjoHLj
— Ben Davis (@davis7) August 21, 2026
As of August 22, there was still no entry for Ox Alpha in Artificial Analysis or LMSys Arena.
This model is available for free, can process approximately 1 million tokens in a single context, and accepts text, image, and video inputs while providing text outputs. It also supports calling tools and functions, although its JSON output lacks schema enforcement, which poses a challenge for developers expecting structured responses.
OpenCode stated that its system can handle 100 trillion tokens daily, while Nous Research, which is providing the model through its own platform, claimed a capacity of 1 quadrillion tokens. Such throughput positions Ox Alpha among the most capable models available today.
These impressive figures suggest that the model is designed for extensive use, and the provider is equipped with sufficient infrastructure to manage the demand.
Ox Alpha (stealth model) is free for the next week
- 1M Context
- Multi-modal
- Zero Data RetentionGenerous rate limits, near unlimited usage
We have capacity for 100T tokens per day, let's see what you can do
— OpenCode (@opencode) August 20, 2026
Reasons for the Buzz
Patrick Collison, CEO of Stripe, whose company announced on August 19 that it would acquire OpenRouter, referred to Ox Alpha as "very impressive" in a post on X, which significantly boosted the model's visibility shortly after its launch.
$ ori --model stealth/ox-alpha
It's very impressive. https://t.co/LRISKwKTOL
— Patrick Collison (@patrickc) August 21, 2026
No one has yet claimed responsibility for the model. AI analyst Andrew Curran noted on Friday that there seems to be uncertainty regarding Ox Alpha's origins.
As the model gained traction, speculation about its creator quickly spread. Developers have suggested possibilities including Microsoft's unreleased MAI family, Xiaomi's MiMo series, DeepSeek, Alibaba's Qwen, and even Google’s Gemini. Some developers humorously remarked that Google might be using a rival's stealth model for its own marketing. However, most of these guesses were dismissed upon closer examination of the evidence.
Xiaomi's MiMo v2.5 incorporates audio input, a feature that Ox Alpha lacks entirely. DeepSeek has not released any video capabilities and provides open weights instead of running stealth previews. Google and Qwen utilize entirely different tokenization and video encoding methods, making it challenging to establish a reliable connection to those models.
What remains points towards Zhipu AI's model family. Independent testing has aligned Ox Alpha's tokenizer with GLM-5.3 across all normalized tests, and its video-token expenditure matched GLM-5V-Turbo precisely across four clips, sharing similar errors and rejection patterns for audio.
Ox Alpha's tokenizer is identical to GLM's. No other lab's tokenizer gets past 4/11 in terms of similarity.https://t.co/3K25G4U2hX
(h/t: @sushsrinivasan)(3/n) pic.twitter.com/dNwwZmCYZ7
— Ananay (@ananayarora) August 21, 2026
However, one detail complicates the narrative: GLM-5.3 was launched as a text-only model on August 14, just six days before Ox Alpha emerged with full video capabilities. If the fingerprinting holds true, it suggests that Ox Alpha might be an unreleased multimodal enhancement to that line, which some analysts have begun calling GLM-5.3 Flash.
Zhipu AI has not made any public comments regarding Ox Alpha's development or origins.
Understanding Stealth Models
A stealth model refers to a high-end AI system released without clear attribution through a routing platform like OpenRouter, allowing a lab to gather real-world data before making a public launch. This approach has become common among Chinese labs. For example, Zhipu AI's Pony Alpha was identified as GLM-5 within five days in February, Xiaomi's Hunter Alpha was revealed as MiMo-V2-Pro in March, and Meituan's Owl Alpha operated anonymously for two months before being confirmed as LongCat-2.0 in June. Ox Alpha is currently available on OpenRouter under the model ID stealth/ox-alpha, as well as through OpenCode, Cline, and Nous Portal, all free to use. OpenCode's announcement indicated that the model would be free for about a week following its August 20 launch, suggesting that the cutoff could be around August 27—after which, following the trend of previous anonymous models, the true identity of the model may finally be revealed.