Ox Alpha, a free "stealth model," launched on August 20 on OpenRouter, OpenCode, Cline, and Nous Research's portal+.

A viral claim that Ox Alpha beats Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on DeepSWE came from a 10-task sample. Two separate full 113-task runs later came in near 63%, roughly in line with GPT-5.6 Sol.

Google, Xiaomi, DeepSeek, and Microsoft's MAI were all floated as possible builders of the model. The strongest theory, however, points to Zhipu AI's GLM-5.3 family.

AI enthusiasts are paying close attention to an AI model called Ox Alpha, which appeared on OpenRouter on August 20 without any company name attached. The listing describes it as "a reasoning model designed for coding, sustained agentic work, and production workloads," built by a third-party provider that has chosen to remain anonymous during the preview.

The mystery model has also drawn notice because it appears to outperform Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on coding benchmarks, a claim that matters because developers are increasingly using these tests to compare models they may not yet be able to identify.

DeepSWE is a benchmark that measures how often a coding agent resolves real GitHub issues correctly on the first attempt, expressed as the percentage of tasks passed across 113 tasks drawn from 91 repositories. On that benchmark, Ox Alpha has been shown to be competitive with top-tier models.

Developer Ben Davis first ran Ox Alpha on a 10-task sample and reported an 80% score on the first pass, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. That result spread quickly and versions of it continued circulating as though Ox Alpha had a clear edge over both models, even though the small sample left room for variance.

I ran this thing through 10 tasks on DeepSWE (so there could be a ton of variance in it's 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 pic.twitter.com/NdDSTjoHLj — Ben Davis (@davis7) August 21, 2026

As of August 22, Ox Alpha still had no listing on Artificial Analysis or Arena, leaving fewer public reference points for users trying to compare it against other frontier models.

The model is free to use, can read up to roughly 1 million tokens in a single context window, and accepts text, image, and video input while returning text. It also supports tool and function calling, although its JSON output is not schema-enforced, which can be a limitation for developers building agents that require structured responses.

OpenCode said its route could handle 100 trillion tokens a day, while Nous Research, which is offering the model free through its own portal, said it had capacity for 1 quadrillion. Tokens are the basic unit of information an AI model processes, and throughput at that scale places the system alongside the most widely used and capable models in the market.

Ox Alpha (stealth model) is free for the next week - 1M Context - Multi-modal - Zero Data Retention Generous rate limits, near unlimited usage We have capacity for 100T tokens per day, lets see what you can do — OpenCode (@opencode) August 20, 2026

Why everyone's paying attention

Stripe CEO Patrick Collison, whose company agreed on August 19 to acquire OpenRouter, called Ox Alpha "very impressive" in a post on X. That comment from a prominent tech executive helped push the model into wider discussion within hours of its launch, especially because the model had already begun circulating through multiple access points before its maker was known.

$ ori --model stealth/ox-alpha

It's very impressive. — Patrick Collison (@patrickc) August 21, 2026

Nobody has publicly claimed responsibility for the model. AI analyst Andrew Curran posted on Friday that people "seem less sure of anything" about where Ox Alpha came from.

As interest grew, speculation quickly broadened. Developers suggested Microsoft's unreleased MAI family, Xiaomi's MiMo line, DeepSeek, Alibaba's Qwen, and even Google's Gemini. So many people landed on Gemini that one developer joked Google was using a rival's stealth model as marketing for its own.

Most of those theories do not fit the evidence. Xiaomi's MiMo v2.5 accepts audio input, a headline feature that Ox Alpha explicitly rejects. DeepSeek has not shipped video capability and releases open weights rather than using stealth previews. Google and Qwen both use different tokenizers and video encoders, making a direct link difficult to support.

The remaining evidence points most strongly to Zhipu AI's models. Independent fingerprinting matched Ox Alpha's tokenizer to GLM-5.3 on every normalized test, and matched its video-token usage to GLM-5V-Turbo exactly across four separate clips. It also shares GLM's distinctive errors and its refusal to accept audio.

Ox Alpha's tokenizer is one for one identical to GLM's. No other lab's tokenizer gets past 4/11 when it comes to tokenizer similarity. (h/t: @sushsrinivasan ) (3/n) pic.twitter.com/dNwwZmCYZ7 — Ananay (@ananayarora) August 21, 2026

One detail, however, complicates a perfect match. GLM-5.3 launched on August 14 as a text-only model, six days before Ox Alpha appeared with full video support. If the fingerprinting is correct, Ox Alpha is more likely an unreleased multimodal upgrade to that line than a direct copy, and analysts have started calling that possibility GLM-5.3 Flash.

Zhipu AI has not publicly commented on Ox Alpha.

What's a stealth model, anyway

A stealth model is a frontier-class AI system released without attribution through a routing platform such as OpenRouter. The arrangement lets a lab gather real-world usage data before committing to a public launch.

The pattern has become common among Chinese labs. Zhipu AI's Pony Alpha was identified as GLM-5 within five days in February. Xiaomi's Hunter Alpha turned out to be MiMo-V2-Pro in March. Meituan kept its Owl Alpha anonymous for two months before confirming it as LongCat-2.0 in June.

Ox Alpha is currently live on OpenRouter under the model ID stealth/ox-alpha, and it is also available through OpenCode, Cline, and Nous Portal, all free to query. OpenCode said the free period runs for about a week from the August 20 debut, which would put the likely cutoff around August 27. After that, if the pattern seen with earlier anonymous models holds, the provider may finally reveal its name.