Ox Alpha: The Stealth Model Beating GPT-5.6 and Grok 4.6
Ox Alpha topped coding and reasoning benchmarks on OpenCode. Here's the evidence pointing to GLM and what it means for builders.
What is Ox Alpha?
Ox Alpha is an anonymous “stealth” language model that appeared on OpenCode with no provider name attached. OpenCode is offering it free for a limited window, paired with a 1 million token context window, multimodal input support, zero data retention, and a claimed daily serving capacity of 100 trillion tokens. Nobody has officially claimed it. But independent benchmark testing and a detailed technical investigation by AI tester Ben Davis point strongly toward GLM, the model family from the Chinese lab Zhipu AI (Z.ai), specifically what could be a next-generation unified multimodal successor to GLM 5.3.
TL;DR
- Ox Alpha appeared unannounced on OpenCode as a free, anonymous model with a 1 million token context window and multimodal support, sparking the usual community guessing game over its origin.
- On Kingbench it scored 87.5% (70 out of 80), placing second overall behind GLM 5.3’s 91.25%, and ahead of Fable 5, Qwen 3.8 Max, and Opus 4.8.
- On a 10-task Deep SWE subset it scored 80%, beating Fable 5 (65%), GLM 5.3 (62%), Grok 4.6 (62%), and GPT 5.6 Sol (52%), and it solved a “Marriott task” in one shot where every other model tested went zero for four.
- Video encoder fingerprinting is the strongest clue to its identity: Ox Alpha’s token counts on controlled video tests matched GLM 5V Turbo exactly, including frame sampling rate, duration scaling near 147 tokens per second, and per-frame resolution scaling.
- Tokenizer counts matched GLM 5.3 across 25 prompts, which effectively requires an identical vocabulary, and the model’s writing style shows the emoji-heavy pattern associated with GLM and Qwen outputs.
- DeepSeek, Qwen, Xiaomi, and Western labs have been ruled out based on these fingerprinting differences, and Davis places roughly 90% confidence on this being a next-gen GLM model, though there is no official confirmation.
- The free access period is time-limited, reportedly ending around August 27, after which the provider may reveal itself the way four previous stealth releases in this pattern have.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How did Ox Alpha perform on benchmarks?
On Kingbench, a coding and reasoning benchmark suite, Ox Alpha scored 70 out of 80 (87.5%), landing in second place on the leaderboard behind GLM 5.3 at 91.25%. That put it ahead of several models generally considered frontier-tier, including Fable 5 (82.5%), Qwen 3.8 Max (81.25%), and Opus 4.8 (80%).
Breaking down individual test results, Ox Alpha earned perfect 10s on a 3JS contact lens rendering task, a Panda SVG generation task, a hard math permutation problem, and a Gemma fine-tuning task. That last one stands out because fine-tuning-related tasks tend to trip up most models. It also scored 8 out of 10 on an elevator simulation and a bow-and-arrow game. Its weakest results were a folding table task and a 3D wrist clock rendering task, both scoring 7. The wrist clock question is notoriously difficult across the board, so a 7 there is actually a strong relative result.
Separately, Ben Davis ran Ox Alpha through a 10-task subset of Deep SWE, an agentic software engineering benchmark, and it scored 80%, with some near-misses suggesting the real number might be slightly higher. That comfortably beat Fable 5 (65%), GLM 5.3 (62%), Grok 4.6 (62%), and GPT 5.6 Sol (52%) on the same subset. Notably, Ox Alpha solved a task referred to as the “Marriott task” in a single attempt, while GLM 5.3, GPT 5.6 Sol, and Grok 4.6 all failed it across four attempts each. Davis cautions that a 10-task subset carries real variance, but a 15-plus point gap over every other tested model is hard to dismiss as noise.
Why do people think Ox Alpha is GLM?
The identification case rests mostly on technical fingerprinting rather than writing style alone, though style plays a supporting role too.
The strongest evidence comes from video encoder behavior. Davis ran controlled video inputs through Ox Alpha and measured how many tokens it consumed under different conditions. The resulting token counts matched GLM 5V Turbo exactly: the same frames-per-second sampling variance, the same duration scaling behavior at roughly 147 tokens per second, and the same way resolution affects token counts per frame. Other providers tested showed distinctly different scaling patterns, which makes this kind of match meaningful rather than coincidental.
Tokenizer behavior added a second, independent data point. Across 25 different prompts, Ox Alpha’s tokenizer produced identical token counts to GLM 5.3. Matching token counts that precisely across a varied prompt set generally requires an identical vocabulary, which is a much stronger signal than surface-level style similarity.
Two smaller signals rounded out the case. Ox Alpha’s responses carry the emoji-decorated formatting style that GLM and Qwen models are known for. And like GLM 5V, it rejects audio input entirely, which rules out models such as MiMo that do support audio.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Using these fingerprints, Davis says he was able to systematically rule out DeepSeek, Qwen, Xiaomi, and Western labs as the source. His conclusion, with roughly 90% confidence, is that Ox Alpha is most likely a next-generation, unified multimodal model from GLM’s maker, not officially released yet. It’s worth noting this is investigative deduction, not a confirmed disclosure, and previous stealth models in this same pattern (this is described as the fifth such release) have all eventually been claimed by a Chinese lab once revealed.
Is Ox Alpha actually stronger than GLM 5.3?
The comparison is more nuanced than a single leaderboard number suggests. On Kingbench, Ox Alpha actually scored slightly below GLM 5.3 (87.5% versus 91.25%). But on the Deep SWE agentic coding subset, it scored well above GLM 5.3 (80% versus 62%).
That split suggests Ox Alpha isn’t simply a re-skinned GLM 5.3. If the GLM attribution holds, it looks more like a distinct checkpoint tuned specifically for agentic, multi-step coding work rather than general one-shot generation and reasoning tasks. That would track with the idea that this is a next-generation model rather than a minor variant or rebrand, since labs training successor models typically shift their tuning priorities based on what real-world usage (like coding agents) demands.
What does the Ox Alpha situation say about the state of AI competition?
Stealth model drops like this have become a recurring pattern: a lab releases a model anonymously through a third-party platform, lets the community benchmark and reverse-engineer it, then reveals itself once the buzz builds. It’s effectively crowdsourced evaluation before a formal launch, and it lets a lab gauge real-world reception without the pressure of an official announcement.
What’s notable here isn’t just the mystery, it’s the performance ceiling. A model widely suspected to be an unreleased checkpoint from a Chinese lab is beating GPT-5.6 Sol and Grok 4.6 on agentic coding tasks, and running competitively with, or ahead of, several models considered frontier-class. Regardless of exactly which lab ships it, that result reinforces a trend that’s been building for a while: the gap between Western frontier labs and Chinese labs on coding and agentic benchmarks has narrowed considerably, and in some specific tasks, reversed.
Frequently Asked Questions
What is Ox Alpha?
Ox Alpha is an anonymous stealth AI model that appeared on OpenCode with no provider name disclosed. It offers a 1 million token context window, multimodal support, zero data retention, and was made free to use for a limited time.
Is Ox Alpha made by GLM?
There’s no official confirmation. Independent testing found that Ox Alpha’s video encoder token counts and tokenizer output match GLM 5V Turbo and GLM 5.3 almost exactly, and other labs were ruled out based on differing technical signatures. The working theory, with an estimated 90% confidence from the investigator who did the fingerprinting, is that it’s an unreleased next-generation GLM model, but this remains speculation until a lab claims it.
How well does Ox Alpha perform on benchmarks?
It scored 87.5% on Kingbench, placing second behind GLM 5.3 and ahead of models like Fable 5, Qwen 3.8 Max, and Opus 4.8. On a 10-task Deep SWE subset, it scored 80%, beating GLM 5.3, Grok 4.6, and GPT 5.6 Sol by a wide margin.
How long will Ox Alpha stay free?
The free access period through OpenCode was reported to run through around August 27, after which pricing or a provider reveal may follow, based on the pattern of previous similar stealth releases.
Why do labs release stealth models instead of announcing them officially?
Releasing a model anonymously lets a lab collect real-world benchmark data and community feedback before committing to an official launch. It also builds anticipation and lets outside testers do informal evaluation work the lab doesn’t have to run itself.
